Traffic volume measurement system and traffic volume measurement method

Abstract

This invention provides a traffic volume measurement system and method that measure traffic volume by determining the direction of vehicle movement at an intersection based on moving images. [Solution] The traffic volume measurement system of the present invention measures traffic volume by determining the direction of travel of a vehicle C based on moving images of a road. The system includes a partitioning means 20 that partitions each still image constituting the moving image into a central zone 21 and a peripheral zone 22 surrounding the central zone, and a travel trajectory derivation means 30 that derives the travel trajectory 51 of the vehicle based on each still image. The travel trajectory derivation means derives the travel trajectory after converting each still image into an image of a specific shape by projection transformation as needed. Furthermore, if the travel trajectory derivation means is unable to derive a part of the travel trajectory and cannot determine the entry or exit position, it estimates the entire travel trajectory based on the travel trajectory that has been derived. By performing projection transformation on each still image, the inability to derive the vehicle's travel trajectory due to adverse conditions such as road shape and the shooting angle of the moving image is reduced.

Description

Technical Field

[0001] The present invention relates to a traffic volume measurement system and a traffic volume measurement method for determining the traveling direction of vehicles at intersections based on moving images and measuring the traffic volume.

Background Art

[0002] There is known a traffic volume measurement system that photographs vehicles passing through a predetermined area such as an intersection and measures the traffic volume of vehicles in each traveling direction based on the video. For example, in the traffic volume measurement device of Patent Document 1, vehicle recognition means recognizes a vehicle based on a video and surrounds the vehicle with a bounding box, virtual passing line setting means sets a virtual passing line so as to intersect the moving direction of the vehicle, and determines whether a predetermined portion of the bounding box has passed through the passing line, thereby measuring the traffic volume. In the traveling direction determination device of Patent Document 2, the side where the vehicle first passes among the four sides constituting the intersection is set as the inflow line, and the side where the vehicle passes among the remaining three sides is set as the outflow line, thereby determining from which direction the vehicle enters the intersection, whether it goes straight, turns right or left and exits, and measuring the traffic volume.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, there is a problem in that it is not easy to accurately determine whether a vehicle has crossed a predetermined line in an image, as is the case with the technology described in the above-mentioned patent document. In addition, when the monitored vehicle crosses other vehicles in an intersection, it momentarily disappears from the image, resulting in problems such as losing sight of the monitored vehicle or misidentifying it as another vehicle. These problems become particularly noticeable when the monitored vehicle turns right or left in an intersection.

[0005] In view of these problems, the present invention aims to provide a traffic volume measurement system and a traffic volume measurement method that measure traffic volume by determining the direction of travel of vehicles at an intersection based on moving images. [Means for solving the problem]

[0006] The traffic volume measurement system of the present invention measures traffic volume by determining the direction of travel of a vehicle on a road based on moving images of the road on which the vehicle is traveling, and comprises: partitioning means for partitioning each still image constituting the moving image into a central zone and a peripheral zone located around the central zone; and trajectory derivation means for deriving the vehicle's trajectory based on each still image, connecting the inflow position where the vehicle flows from the peripheral zone into the central zone and the outflow position where the vehicle flows out to the peripheral zone, wherein the trajectory derivation means derives the trajectory after converting each still image into an image of a specific shape by projection transformation as necessary, and further, if the trajectory derivation means cannot derive a part of the trajectory and cannot determine at least one of the inflow position and the outflow position, it is characterized in that it estimates the entire trajectory, including the undetermined part of the trajectory, based on the trajectory that has been derived. Furthermore, the system is characterized by providing data display means that displays an image of each vehicle, showing at least one of the following items: an image of the vehicle, the type of vehicle, the direction of travel of the vehicle, the time when the vehicle reached the entry and exit positions, the distance of the travel trajectory, a diagram of the travel trajectory, and whether the travel trajectory was derived using the "zoning determination method" or the "AI determination method". Furthermore, the system is characterized by allowing the user to visually confirm each of the items displayed in the image and correct them as needed.

[0007] The present invention provides a traffic volume measurement method for determining the direction of travel of a vehicle on a road based on moving images of the road on which the vehicle is traveling, and measuring the traffic volume, comprising the steps of: dividing each still image constituting the moving image into a central zone and a peripheral zone located around the central zone; deriving a vehicle's trajectory based on each still image, connecting an inflow position where the vehicle flows from the peripheral zone into the central zone and an outflow position where the vehicle flows out to the peripheral zone; deriving the trajectory after converting each still image into an image of a specific shape by projection transformation as necessary; and, if a part of the trajectory cannot be derived and at least one of the inflow position and the outflow position cannot be determined, estimating the entire trajectory, including the undetermined part of the trajectory, based on the derived trajectory. Furthermore, the system is characterized by providing a step of displaying an image of each vehicle, showing at least one of the following items: an image of the vehicle, the vehicle type, the direction of travel of the vehicle, the time when the vehicle reached the entry and exit positions, the distance of the travel trajectory, a diagram of the travel trajectory, and whether the travel trajectory was derived using the "zoning determination method" or the "AI determination method". Furthermore, the system is characterized by including a step that allows the user to visually confirm each of the items displayed in the image and correct them as necessary. [Effects of the Invention]

[0008] In this invention, the vehicle's trajectory is made easier to derive by applying a projection transformation to each still image that makes up the moving image as needed. Furthermore, if the monitored vehicle is lost sight of and part of its trajectory cannot be determined, the entire trajectory is estimated based on the trajectory that could be determined. Therefore, the direction of travel of the vehicle can be accurately determined. [Brief explanation of the drawing]

[0009] [Figure 1] Block diagram showing the configuration of the traffic volume measurement system according to the first embodiment of the present invention [Figure 2] Diagram showing a state in which a still image is partitioned into a central zone and a peripheral zone by partitioning means [Figure 3] Figures (a) to (d) for explaining projective transformation [Figure 4] Figures (a) to (e) for explaining the method of projective transformation [Figure 5] Figures (a) and (b) for explaining the method of projective transformation [Figure 6] Figures (a) and (b) for explaining the method of projective transformation [Figure 7] Figures (a) showing a table for classifying the traveling direction and (b) showing the traveling direction of a vehicle [Figure 8] Diagram showing a state in which a part of the traveling locus is interrupted [Figure 9] Figures (a) to (c) for explaining the method of machine learning [Figure 10] Flowchart showing an example of each process of the traffic volume measurement method [Figure 11] Flowchart showing an example of each process of the traffic volume measurement method [Figure 12] Flowchart showing an example of each process of the traffic volume measurement method [Figure 13] Block diagram showing the configuration of the traffic volume measurement system according to the second embodiment [Figure 14] An example of image display by data display means [Figure 15] An example of screen display when modifying each item [Figure 16] Report screen of the analyzed and corrected traffic volume measurement result [Embodiments for Carrying Out the Invention]

[0010] [First Embodiment] The first embodiment of the traffic volume measurement system and traffic volume measurement method of the present invention will be described. The traffic volume measurement system is a system for determining the traveling direction of vehicles on a road and measuring the traffic volume based on a moving image of the road. The moving image may be recorded on an information storage medium after shooting, or may be a live moving image. In this specification, "road" includes not only a single straight road but also intersections, and "intersection" includes crossroads, T-junctions, and other locations where two or more roads intersect. That is, the traffic volume measurement system of the present invention can determine the traveling direction of vehicles passing through a predetermined area on a road such as an intersection and measure the traffic volume.

[0011] As shown in FIG. 1, the traffic volume measurement system 1 is generally composed of a control unit 10, a partitioning means 20, a traveling trajectory derivation means 30, a storage unit 40, etc. Although not shown in the figure, the control unit 10 includes a CPU, a RAM, and a ROM. The CPU reads various programs stored in the ROM and various information stored in the storage unit 40 and executes them as appropriate to comprehensively control the traffic volume measurement system 1.

[0012] As shown in FIG. 2, the partitioning means 20 performs a process of partitioning each still image (frame) constituting the moving image into a central zone 21 and a peripheral zone 22 located around the central zone 21. In the present embodiment, the intersection 50 is taken as the central zone 21, and the periphery of the intersection 50 is taken as the peripheral zone 22. Specifically, the partitioning means 20 partitions the zone on the left side of the central zone 21 in the peripheral zone 22 as the A zone 22a, the upper side as the B zone 22b, the right side as the C zone 22c, and the lower side as the D zone 22d. The traveling trajectory derivation means 30 derives a traveling trajectory that connects the inflow position and the outflow position of a vehicle from the peripheral zone 22 (A to D zones 22a to 22d) to the central zone 21 and then back to the peripheral zone 22 (A to D zones 22a to 22d) based on each still image. The traveling trajectory derivation means 30 derives the traveling trajectory of the vehicle using a well-known object detection algorithm such as YOLO (You Only Look Once) or a well-known object tracking algorithm such as BoT-SORT (Robust Associations Multi-Pedestrian Tracking).

[0013] The trajectory derivation means 30 performs a projection transformation when deriving the trajectory of the vehicle. Figure 3 shows an example of a projection transformation. Figure 3(a) schematically represents a still image that makes up a video of the intersection 50. A video camera was installed on a high point at one corner of the intersection 50 to capture the entire intersection 50, and the intersection 50 was filmed from a diagonal downward angle, so the intersection 50 appears as a distorted rectangle. It is not easy for the trajectory derivation means 30 to derive the trajectory of vehicle C based on this still image. This is because, as shown in Figures 3(b) and (c), when vehicle C passes at a position far from the video camera within the intersection 50, its trajectory 51 appears small and distorted. Therefore, as shown in Figure 3(d), the distorted rectangle capturing the intersection 50 is transformed into a square by projection transformation to make it easier to derive the vehicle's trajectory 51.

[0014] In practice, multiple still images arranged in a time series as shown in Figures 4(a) to 4(e) are combined to create a single still image depicting the vehicle's trajectory 51 within the intersection 50, as shown in Figure 5(a). This image is then transformed into a square using a projection transformation, as shown in Figure 5(b). Alternatively, the intersection in each still image can be transformed into a square using a projection transformation, and these transformed still images can be combined to create a single square still image depicting the vehicle's trajectory 51 within the intersection 50. Furthermore, the shape of the image after the projection transformation does not necessarily have to be a square; for example, it may be a rectangle, ellipse, circle, etc.

[0015] Figure 5(a) is an image of the trajectory 51 before projection transformation when a vehicle enters the intersection 50 of the central zone 21 from zone A 22a, turns left, and exits into zone B 22b. As described above, the video camera is positioned diagonally downwards from a high vantage point, and the vehicle's trajectory 51 within the intersection 50 approximates a single line extending horizontally. Therefore, it is difficult for the trajectory derivation means 30 to accurately derive the trajectory 51 based on this image. Therefore, by transforming the intersection 50 into a square as shown in Figure 5(b) using projection transformation, the travel trajectory 51 is transformed into a line that curves upward from the left center, so the travel trajectory derivation means 30 can deduce that the vehicle entered the intersection 50 in the central zone 21 from zone A 22a, turned left, and exited into zone B 22b (third step). In Figures 5(a) and 5(b), it is not possible to determine which of the two ends of the trajectory 51 represents the entry point (the position where the vehicle entered the central zone 21) and which represents the exit point (the position where the vehicle exited the central zone 21). Therefore, it would be helpful to indicate the entry point with a conspicuous color such as red.

[0016] Similarly, Figure 6(a) is an image of the trajectory 51 before projection transformation when a vehicle flows from zone D 22d into intersection 50 in central zone 21, turns left, and exits into zone A 22a. The trajectory 51 appears as a single unit, making it difficult for the trajectory derivation means 30 to accurately derive the vehicle's trajectory 51 based on this image. Therefore, by transforming the intersection 50 into a square as shown in Figure 6(b) through projection transformation, the travel trajectory 51 is transformed into a line that curves to the left from the bottom upwards, so the travel trajectory derivation means 30 can deduce that the vehicle entered the intersection 50 in the central zone 21 from zone D 22d, turned left, and exited into zone A 22a.

[0017] Figure 7(a) is a table showing 12 patterns of travel direction classifications, indicating which surrounding zone 22 (zones A-D 22a-22d) a vehicle enters the central zone 21 from, proceeds straight, turns right or left in the central zone 21, and exits to which surrounding zone 22 (zones A-D 22a-22d). Figure 7(b) shows the direction of travel of the vehicle. For example, direction 1 in the direction of travel classification table represents a pattern where a vehicle flows from zone A 22a into central zone 21 and exits into zone B 22b. In this case, the vehicle would have turned left in central zone 21. For example, direction 5 represents a pattern in which a vehicle flows from zone B 22b into the central zone 21 and out into zone D 22d, in which case the vehicle traveled straight through the central zone 21.

[0018] By counting the number of vehicles in each direction (1-12) from the start to the end of a video, it is possible to measure traffic volume by direction of travel. Furthermore, by considering the time each still image was taken, it is also possible to measure traffic volume by time of day. The above describes a traffic volume measurement method (referred to as the "zoning determination method" in this specification) when the entire trajectory 51 can be derived by the trajectory deriving means 30 being able to identify the entry and exit positions of the vehicle. Furthermore, the trajectory derivation means 30 does not necessarily need to perform projection transformation when deriving the vehicle's trajectory. In other words, for example, in the state shown in Figure 3(c), if vehicle C passes close to the video camera in the intersection 50, its trajectory 51 will be clearly and largely visible, so the trajectory derivation means 50 may derive the trajectory 51 without performing projection transformation. In this case, the time required for projection transformation is eliminated, and the time required for traffic volume measurement can be shortened. Thus, the trajectory derivation means 30 transforms each still image into an image of a specific shape by projection transformation as needed.

[0019] Next, we will describe a traffic volume measurement method (referred to as the "AI determination method" in this specification) for cases where the trajectory derivation means 30 is unable to derive a portion of the trajectory 51 and is unable to determine at least one of the vehicle's entry and exit positions. In this case, the trajectory derivation means 30 estimates the entire trajectory 51 of the vehicle within the central zone 21, including some of the trajectory that could not be derived, based on the trajectory 51 that has been derived. For example, a vehicle being tracked may momentarily disappear from the still image as it crosses another vehicle at intersection 50, resulting in the vehicle being lost from view. In this case, the trajectory derivation means 30 cannot fit the vehicle's direction of travel into any of the 12 patterns mentioned above. Therefore, it uses the portion of the trajectory 51 that has been derived to estimate the entire trajectory 51, including the portion that could not be derived (the missing portion).

[0020] For example, as shown in Figure 8, when we look at the trajectory 51 after projection transformation, we can see that the vehicle entered the intersection 50 of the central zone 21 from zone B 22b, but because the trajectory 51 is interrupted, we may not be able to determine which surrounding zone 22 it exited to. In this case, the trajectory derivation means 30 estimates the entire trajectory 51 based on the derived trajectory 51. Machine learning such as a Convolutional Neural Network (CNN) is used as a means to have the trajectory derivation means 30 estimate the entire trajectory 51. Specifically, as shown in Figure 9(a), a large number of still image data are prepared in which the trajectory 51 is drawn to the extent that the inflow and outflow positions can be identified, before projection transformation. Next, as shown in Figure 9(b), a certain percentage (e.g., 40%) is deleted from the end of the trajectory 51 for each still image, and as shown in Figure 9(c), the images obtained by projection transformation are used as training data. The data of the trajectory 51 from which a certain percentage has been deleted from the training data corresponds to the input data, and the correct set of inflow and outflow positions corresponds to the correct label. The training data is then appropriately divided and used as training data, validation data, and test data. If the amount of data is insufficient, methods such as cross-validation may be used to efficiently utilize the data.

[0021] Since machine learning methods are well-known, a detailed explanation will be omitted, but machine learning is completed through steps such as algorithm selection, model training using training data, model evaluation using validation data, and model testing using test data. In the example above, training data was created by deleting a certain percentage (40%) from the end of the trajectory 51 for each still image. However, training data with deletion rates of 0%, 20%, 40%, and 80% could also be used, and the optimal deletion rate could be determined based on the results of model testing. The optimal model selected by machine learning is stored in the memory unit 40, and the trajectory derivation means 30 can use the optimal model to estimate the entire trajectory 51 if it is unable to determine the exit position.

[0022] Furthermore, by deleting a certain percentage (for example, 40%) from the beginning of the trajectory 51 for each still image, and using the resulting projected image as training data for machine learning and model improvement, it becomes possible to estimate the entire trajectory 51 using the optimal model when the trajectory derivation means 30 is unable to identify the inflow position. Furthermore, by deleting a certain percentage from the start and a certain percentage from the end of the trajectory 51, and using the resulting projected image as training data for machine learning and model improvement, it becomes possible to estimate the entire trajectory 51 using the optimal model when the trajectory derivation means 30 is unable to identify both the inflow and outflow positions. Alternatively, a large number of projected still image data are prepared in which the trajectory 51 is drawn to the extent that the inflow and outflow positions can be identified. For each still image, a certain percentage (for example, 40%) of the trajectory 51 is removed, and these images are used as training data.

[0023] Figure 10 is a flowchart showing an example of each process in a traffic volume measurement method. First, the partitioning means 20 performs a process to partition each still image that makes up the moving image into a central zone 21 and a peripheral zone 22 located around the central zone 21 (step S1). Next, the trajectory derivation means 30 converts each still image into an image of a specific shape by projection transformation (step S2). Next, the trajectory derivation means 30 derives a trajectory 51 of the vehicle from the surrounding zone 22 to the central zone 21 and then back to the surrounding zone 22 based on each still image after the projection transformation (YES in step S3).

[0024] If the answer in step S3 is YES, the progress trajectory derivation means 30 selects the appropriate progress trajectory 51 from among the 12 patterns based on the progress classification table, stores it in the storage unit 40 (step S4), and the process ends. If the answer in step S3 is NO, that is, if a portion of the trajectory 51 could not be derived, the trajectory derivation means 30 estimates the entire trajectory 51 based on the derived trajectory 51 (step S5). Specifically, it estimates the entire trajectory 51 using the optimal model selected by machine learning, and the process ends after step S4.

[0025] Alternatively, as shown in the flowchart of Figure 11, the trajectory derivation means 30 may derive the trajectory without performing a projection transformation (Yes in step S11). If the trajectory derivation means 30 is unable to derive part of the trajectory (No in step S11), the entire trajectory 51 may be estimated based on the derived trajectory 51 (step S12), and then the entire trajectory may be derived after performing a projection transformation (step S13) (step S14). Furthermore, as shown in the flowchart of Figure 12, the entire trajectory may always be estimated (step S21), regardless of whether the trajectory derivation means 30 can derive the trajectory or not, or whether it can derive only a part of it. In this case, the projection transformation may be performed before the step of estimating the entire trajectory (step S21).

[0026] Next, a second embodiment of the traffic volume measurement system of the present invention will be described, but parts that have the same configuration as in the first embodiment will be denoted by the same reference numerals and their descriptions will be omitted. This embodiment is characterized by the inclusion of a data display means 60, as shown in Figure 13. The data display means 60 displays at least one of the following items for each vehicle: an image of the vehicle, the vehicle type, the direction of travel of the vehicle, the time the vehicle reached the entry and exit positions, the distance of the travel trajectory, a diagram of the travel trajectory, and whether the travel trajectory was derived using the "zoning determination method" or the "AI determination method". Figure 14 is an example of image display by the data display means 60, showing the vehicle image 61, vehicle type 62 (small car or large car), vehicle direction 63, time 64 when the vehicle reached the entry point, distance 65 of the trajectory, diagram 66 of the trajectory, and determination method 67 ("zoning determination method" or "AI determination method"). In Figure 14, the intersection is defined as the central zone, and the upper part of the central zone is divided into Zone A, the right side into Zone B, the lower part into Zone C, and the left side into Zone D. Furthermore, as shown in Figure 15, the user may be able to visually confirm each item displayed in the image and make corrections as needed. The analyzed and corrected traffic volume measurement results are displayed as a report screen, as shown in Figure 16, and can be downloaded as needed. The analyzed and corrected traffic volume measurement results can also be used as data for the machine learning described above. [Industrial applicability]

[0027] The present invention relates to a traffic volume measurement system and method that measure traffic volume by determining the direction of travel of vehicles at an intersection based on moving images, and has industrial applicability. [Explanation of symbols]

[0028] C Vehicle 1. Traffic volume measurement system 10 Control Unit 20 partitioning means 21 Central Zone 22 Surrounding Zones 22a Zone A 22b Zone B 22c C zone 22d D zone 30 Progress trajectory derivation means 40 Storage section 50 Intersections 51 Progress trajectory 60 Data display means 61 Vehicle images 62 car models 63. Direction of travel of the vehicle 64 Time when the vehicle reached the entry point 65 Distance of the trajectory 66 Diagram of the trajectory 67 Judgment method

Claims

1. In a traffic volume measurement system that measures traffic volume by determining the direction of travel of a vehicle on a road based on video footage of the road on which the vehicle is traveling, A partitioning means that performs a process to divide each still image constituting the moving image into a central zone and a peripheral zone located around the central zone, The system includes a trajectory derivation means for deriving a vehicle's trajectory that connects an inflow position where the vehicle flows from the surrounding zone to the central zone and an outflow position where the vehicle flows out to the surrounding zone, based on each of the aforementioned still images. The aforementioned trajectory derivation means derives the trajectory by converting each of the still images into an image of a specific shape by projection transformation as necessary. Furthermore, the traffic volume measurement system is characterized in that, if the trajectory derivation means is unable to derive a portion of the trajectory and cannot determine at least one of the inflow position and the outflow position, it estimates the entire trajectory, including the portion of the trajectory that could not be derived, based on the trajectory that could be derived.

2. Furthermore, the traffic volume measurement system according to claim 1 is characterized by comprising data display means that displays an image of each vehicle, the vehicle type, the direction of travel of the vehicle, the time when the vehicle reached the entry position and the exit position, the distance of the travel trajectory, a diagram of the travel trajectory, and whether the derivation of the travel trajectory was performed using the "zoning determination method" or the "AI determination method".

3. Furthermore, the traffic volume measurement system according to claim 2 is characterized in that the user can visually confirm each of the items displayed in the image and correct them as necessary.

4. In a traffic volume measurement method that determines the direction of travel of a vehicle on a road based on video footage of the road on which the vehicle is traveling, The process involves dividing each still image constituting the moving image into a central zone and peripheral zones located around the central zone. The steps include: deriving the vehicle's trajectory based on each of the aforementioned still images, connecting the inflow position where the vehicle flows from the surrounding zone to the central zone and the outflow position where it flows out to the surrounding zone; The steps include: deriving the trajectory by converting each of the aforementioned still images into an image of a specific shape by projection transformation as necessary; A traffic volume measurement method characterized by comprising the step of estimating the entire trajectory, including the portion of the trajectory that could not be derived, based on the derived trajectory, when a portion of the trajectory cannot be derived and at least one of the inflow position and the outflow position cannot be determined.

5. Furthermore, the traffic volume measurement method according to claim 4 is characterized by further comprising the step of displaying an image of each vehicle, including at least one of the following items: an image of the vehicle, the type of vehicle, the direction of travel of the vehicle, the time when the vehicle reached the entry position and the exit position, the distance of the travel trajectory, a diagram of the travel trajectory, and whether the derivation of the travel trajectory was performed using the "zoning determination method" or the "AI determination method".

6. Furthermore, the traffic volume measurement method according to claim 5 is characterized by comprising a step in which the user can visually confirm each of the items displayed in the image and correct them as necessary.

Images