Moving object detection system and moving object detection method

Abstract

To improve a detection accuracy of a moving object in a moving object detection system without incurring a long-term measurement or incurring a cost.SOLUTION: A moving object detection system according to the present invention reads vibration data as learning-purpose measuring data, and generates two or more pieces of learning-purpose data from the learning-purpose measuring data, some of which are used for forming a learning model that overlaps in time with the other learning-purpose data. Then, the moving object detection system forms the learning model based on the learning-purpose data.SELECTED DRAWING: Figure 2

Description

[Technical field]

[0001] The present invention relates to a moving object detection system and a moving object detection method for detecting the movement of a moving object. [Background technology]

[0002] The applicant has filed an invention for detecting the movement of a moving body such as a vehicle using a vibration sensor (see Patent Document 1). When the movement of a moving body is detected using a vibration sensor as in the invention of this patent application, for example, the number of passing (moving) vehicles, which are moving bodies, is detected. In this case, for example, a learning model for detecting the number of passing vehicles is created. That is, first, the vibration sensor measures (acquires) the vibration of the road caused by the passing of the vehicle. Next, peak detection is performed on the vibration data of the road caused by the passing of the vehicle measured (acquired) by the vibration sensor (measurement data for learning, vibration data for learning), and a learning model in which the peak part is learned is created. In addition, the vibration of the road caused by the passing of the vehicle that is the target for detecting the number of passing vehicles is measured. Then, the number of passing vehicles is detected based on the created learning model and the measured vibration data of the road caused by the passing of the vehicle (measurement data).

[0003] In such a moving object detection system, if a learning model with an insufficient number of learning data is created, the created learning model will be a model that follows the tendency of the learning data. As a result, the detection accuracy is good for learned data, but for unknown data, overlearning, in which the detection accuracy decreases, becomes a problem. For example, assume that a learning model is created using only road vibration data caused by the passage of a vehicle with a slow speed as learning data. In this case, if the measured road vibration data is data caused by the passage of a vehicle with a slow speed, a vehicle can be detected based on the learning model. However, if the measured road vibration data is data caused by the passage of a vehicle with a fast speed that has different characteristics from the road vibration data caused by the passage of a vehicle with a slow speed, a problem occurs in that the vehicle cannot be detected. Therefore, in order to prevent overlearning, a sufficient number of learning data must be acquired. However, in the conventional method, learning data is created by simply dividing the learning measurement data, so a long measurement time is required to acquire a sufficient number of learning data to prevent overlearning.

[0004] Also, for example, when detecting the number of pedestrians, there is a technique for detecting the number of pedestrians using four vibration sensors (see Non-Patent Document 1). However, the method using multiple sensors has a problem of high cost because multiple sensors are required. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Patent Publication No. 2022-022822 [Non-patent literature]

[0006] [Non-Patent Document 1] Yu Yang, 4 others, “Pedestrian Counting Based on Piezoelectric Vibration Sensor”, [online], February 12, 2022, MDPI, [searched on May 15, 2023],<URL:https: / / www.mdpi.com / 2076-3417 / 12 / 4 / 1920> Summary of the Invention [Problem to be solved by the invention]

[0007] As described above, sufficient learning data is required to create a learning model with high detection accuracy that does not cause overlearning. However, there are various problems such as the need for long-term measurements to obtain a large amount of learning data and the need for multiple sensors, which is costly.

[0008] . An object of the present invention is to improve the accuracy of detecting a moving object without performing long-term measurements or incurring costs. [Means for solving the problem]

[0009] The moving object detection system of the first invention is characterized in that it creates two or more pieces of learning data for creating a learning model from learning measurement data, which is measurement data of a moving object, some of which overlaps in time with other learning data, creates the learning model for detecting a moving object based on the learning data, and detects a moving object based on the created learning model and the measurement data of the moving object.

[0010] In conventional methods, learning data was created by simply dividing the learning measurement data, so the number of learning data was small, and overlearning occurred when a learning model was created from a small amount of data, in which the learned data was detected accurately but the detection accuracy for unknown data was reduced. The moving object detection system of the present invention creates two or more pieces of learning data from the learning measurement data to create a learning model in which some of the learning data overlaps in time with other learning data. The moving object detection system then creates the learning model based on the learning data. In this way, the moving object detection system of the present invention can increase the number of learning data compared to conventional methods by creating learning data that overlaps in time.

[0011] Therefore, a learning model can be created with a sufficient amount of data, and overfitting can be prevented. By detecting a moving object using a learning model in which overfitting has been prevented, it is possible to improve the accuracy of detecting a moving object. In addition, since the amount of learning data can be increased without using multiple sensors, the number of sensors can be reduced, thereby reducing costs. In addition, since the amount of learning data required to improve accuracy can be obtained without measuring, the measurement time required to obtain data can be reduced.

[0012] That is, according to the present invention, it is possible to improve the detection accuracy of a moving object without performing long-term measurements or incurring costs.

[0013] The moving body detection system of the second invention is characterized in that, in the moving body detection system of the first invention, the learning data is created by sliding the learning measurement data along the time axis by a predetermined time at a time and extracting a predetermined length of data from the slid time.

[0014] For example, when creating 30-second pieces of learning data from 60 seconds of learning measurement data, 61 pieces of learning data can be created by sliding the data along the time axis by 0.5 seconds at a time.

[0015] The moving body detection system of the third invention is characterized in that, in the moving body detection system of the first invention, the learning data is created by extracting data of a predetermined length from the learning measurement data, some of which overlaps in time with other learning data.

[0016] A moving object detection system according to a fourth aspect of the present invention is the moving object detection system according to the first aspect of the present invention, characterized in that the learning measurement data and the measurement data are vibration data.

[0017] A moving object detection system according to a fifth aspect of the present invention is the moving object detection system according to the first aspect of the present invention, characterized in that the moving object measures vibrations generated in a moving target on which the moving object moves, and the measured vibration data is used as the learning measurement data.

[0018] The moving body detection system of the sixth invention is characterized in that, in the moving body detection system of the first invention, vibrations occurring in a moving target along which the moving body moves are measured, vibration peaks are extracted from the measured vibration data, and the extracted vibration peaks are used as the learning measurement data.

[0019] A moving object detection system according to a seventh aspect of the present invention is the moving object detection system according to the first aspect of the present invention, characterized in that the number of moving objects is detected based on the created learning model and measurement data.

[0020] A moving object detection system according to an eighth aspect of the present invention is the moving object detection system according to the first aspect of the present invention, and detects the type of the moving object based on the created learning model and measurement data.

[0021] A moving object detection system according to a ninth aspect of the present invention is the moving object detection system according to the first aspect of the present invention, characterized in that the moving object is a pedestrian.

[0022] A moving object detection system according to a tenth aspect of the present invention is the moving object detection system according to the first aspect of the present invention, characterized in that the moving object is a vehicle.

[0023] The moving body detection system of the 11th invention comprises a sensor and a control unit, and the control unit creates two or more pieces of learning data from learning measurement data, which is measurement data of a moving body measured by the sensor, to create a learning model, some of which overlaps in time with other learning data, creates the learning model for detecting a moving body based on the learning data, and detects the moving body based on the created learning model and the measurement data measuring the moving body.

[0024] The moving object detection system of the 12th invention is characterized in that it creates two or more pieces of learning data from learning measurement data, which is measurement data of a moving object, to create a learning model in which some of the learning data overlaps in time with other learning data, creates the learning model for detecting a moving object based on the learning data, and detects the moving object based on the created learning model and the measurement data measuring the moving object. Effect of the Invention

[0025] According to the present invention, it is possible to improve the detection accuracy of a moving object without performing long-term measurements or incurring high costs. [Brief description of the drawings]

[0026] [Figure 1] 11 is a diagram for explaining a process of creating learning data based on learning measurement data. FIG. [Diagram 2] 10 is a flowchart showing a processing operation of the moving object detection system for creating learning data. [Diagram 3] 5 is a flowchart showing a learning process operation of the moving object detection system according to the present embodiment. [Figure 4] 5 is a flowchart showing a vehicle detection processing operation of the moving object detection system according to the present embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0027] Hereinafter, an embodiment of the present invention will be described. A moving body detection system according to an embodiment of the present invention detects the movement of a moving body such as a vehicle, and counts the number of detected moving bodies. The moving body detection system includes a vibration sensor, a control unit such as a CPU (Central Processing Unit), and the like. The moving body detection system may be composed of one device, or may be composed of two or more devices. In addition, the detection of a moving body based on vibration data detected by the vibration sensor may be performed by, for example, a server (cloud). Note that the processes and the like described below are basically executed by the control unit, but in the process of "the control unit...", the wording "the control unit" may be omitted. Hereinafter, an embodiment will be described in which the moving body is a vehicle, and the number of vehicles passing through a road that is a moving target is counted.

[0028] A moving object detection system according to an embodiment of the present invention detects vehicle movement from vibration data output from a vibration sensor using a learning model created by machine learning, and counts the number of detected vehicles. Machine learning is a technology for making a computer acquire human-like learning capabilities, and refers to a technology in which a computer creates a learning model from learning data that is previously acquired, and applies this model to new data to perform detection.

[0029] The learning model refers to a model that is trained in advance using appropriate learning data for a machine learning model according to an arbitrary machine learning algorithm such as deep learning. In this embodiment, a learning model for detecting a vehicle is created using learning data based on vibration data (learning measurement data) for creating a learning model output from a vibration sensor. Note that a regression model is adopted as the learning model in this embodiment. The regression model is a model that is constructed so that a machine learning algorithm can learn using teacher data and predict numerical values.

[0030] The vibration sensor is installed, for example, on a road or the like. The vibration sensor detects vibrations transmitted to the road from a vehicle passing on the road and outputs vibration data. The control unit creates learning data based on the vibration data (learning measurement data) output from the vibration sensor.

[0031] Fig. 1 is a diagram for explaining a process of creating learning data based on learning measurement data. Fig. 1(a) is a diagram showing learning measurement data. The learning measurement data is vibration data obtained by detecting vibrations that occur on a road when a vehicle passes along the road using a vibration sensor. In this embodiment, two or more pieces of learning data are created from the learning measurement data, which is measurement data of a vehicle, with some of the pieces of learning data overlapping in time with other pieces of learning data.

[0032] When creating learning data, the control unit first extracts a predetermined length of data from the measurement start time (0 seconds) of the learning measurement data. In this embodiment, the predetermined length is 30 seconds, and the control unit extracts 30 seconds of data from the measurement start time of the learning measurement data, as shown in FIG. 1(b-1). The extracted 30 seconds of data is used as learning data. The time at which extraction of data as the first learning data begins (hereinafter referred to as the "extraction start time") is the measurement start time (0 seconds) of the learning measurement data.

[0033] Next, the control unit slides the extraction start time in the time axis direction by a predetermined time and extracts data of a predetermined length. In this embodiment, the predetermined time is 0.5 seconds, and the control unit extracts data of 30 seconds from the extraction start time (0.5 seconds) that is slid 0.5 seconds from the measurement start time in the time axis direction, as shown in FIG. 1(b-2). Next, the control unit extracts data of 30 seconds from the extraction start time (1.0 seconds) that is slid 0.5 seconds further in the time axis direction from the point 0.5 seconds after the measurement start time, as shown in FIG. 2(b-3). In this way, the control unit creates learning data by sliding the extraction start time in the time axis direction by a predetermined time (0.5 seconds) increments in the learning measurement data and extracting data of a predetermined length (30 seconds) from the slid time.

[0034] The processing operation of the moving object detection system for creating learning data will be described with reference to the flowchart shown in FIG. 2. The control unit cuts out vibration data to a preset length (60 seconds) and reads it as learning measurement data (S1). Next, the control unit extracts data of a predetermined length (30 seconds) from the learning measurement data to create learning data (S2). Next, the control unit judges whether the time at which the extraction of the extracted learning data ends (hereinafter, "extraction end time") matches the measurement end time of the learning measurement data (S3). If the control unit judges that the extraction end time of the extracted learning data does not match the measurement end time of the learning measurement data (S3: NO), it slides the extraction start time of the learning data in the time axis direction by a predetermined time (0.5 seconds) (S4) and executes the processing of S2.

[0035] When the control unit determines that the extraction end time of the extracted learning data matches the measurement end time of the learning measurement data (S3: YES), the control unit ends the process. As described with reference to FIG. 1, the control unit repeats sliding in the time axis direction until the extraction end time of the learning data matches the measurement end time of the learning measurement data, and creates multiple pieces of learning data.

[0036] For example, assume that the length of the learning measurement data is 60 seconds. The control unit slides the extraction start time by 0.5 seconds until the extraction end time of the learning data with a predetermined length of 30 seconds from the 60-second vibration data matches the measurement end time of the learning measurement data. In this case, if the extraction start time is slid 60 times by 0.5 seconds, the end time of the learning data (30 seconds + 0.5 seconds x 60 times = 60 seconds) matches the measurement end time of the learning measurement data (60 seconds), so that 61 pieces of learning data are created. Note that, if the sliding time is 0.5 seconds, the length of the learning measurement data is set to a length divisible by 0.5 seconds, such as 60 seconds, so that in S3, it is determined that the extraction end time of the extracted learning data matches the measurement end time of the learning measurement data.

[0037] Next, the learning process operation of the moving object detection system according to this embodiment will be described with reference to the flowchart shown in Fig. 3. The control unit creates learning data using the above-mentioned method (S21). Next, the control unit counts the number of passing vehicles in the time period corresponding to each created learning data, and creates a correct answer label (S22). For example, the control unit uses video recognition for video data capturing a moving vehicle to recognize the vehicle's passing time, and creates a correct answer label that records the vehicle's passing time.

[0038] Next, the control unit extracts features using a log-Mel-spectrogram based on the learning data (S23).

[0039] A log-mel spectrogram performs frequency analysis continuously over time, and uses color to represent intensity, providing a three-dimensional display of intensity, frequency, and time. In addition, a mel spectrogram displays the frequency axis in a frequency scale called the mel scale, in order to bring the frequency spectrum of a signal closer to the characteristics of human hearing. The mel scale is designed to represent frequency changes more finely in low frequency bands and more gradually in high frequency bands. A log-mel spectrogram performs a logarithmic operation on a mel spectrogram, which makes it possible to display the intensity and amplitude of a signal logarithmically.

[0040] In this embodiment, the feature amount is extracted using a log-mel spectrogram. However, the feature amount may be extracted by other methods such as Mel Frequency Cepstrum Coefficients (MFCC).

[0041] Next, the control unit associates the features extracted in S23 with the time period of the correct label for the number of passing vehicles created in S22, and trains them to create a learning model using deep learning (S24).

[0042] Deep learning is one of the elemental technologies of AI (artificial intelligence). Deep learning is learning using deep neural networks that allows machines to automatically extract features from data without human intervention, provided there is a sufficient amount of data. Deep learning increases information transmission and processing by creating multiple intermediate layers between the input layer and output layer, thereby improving the accuracy and versatility of features and detection accuracy.

[0043] Next, the vehicle detection processing operation of the moving object detection system according to this embodiment will be described with reference to the flowchart shown in FIG. 4. The control unit reads vibration data output from the vibration sensor (S31). Next, the control unit divides the read vibration data into data of a predetermined length (30 seconds) (S32). Since the learning model in this embodiment is created using learning data with a predetermined length of 30 seconds, in order to detect the number of passing vehicles, it is necessary to match the vibration data to the length of the learning data. For this reason, the vibration data is divided every 30 seconds.

[0044] Next, the control unit extracts features from each divided data based on the divided data (S33). The control unit extracts features using the same extraction method as that used to create the learning model. That is, in this embodiment, the feature is extracted using a log-mel spectrogram. Next, the control unit detects the number of passing vehicles based on the learning model and the feature of the data (S34). Next, the control unit counts the number of passing vehicles detected from each divided data (S35). In this embodiment, the number of passing vehicles is detected every 30 seconds based on a learning model with a length of 30 seconds. That is, the number of passing vehicles, such as 2, 3, 5, etc., is detected every 30 seconds. Then, these numbers of passing vehicles are counted.

[0045] Based on the learning data created by the above-mentioned method, learning is performed and a learning model is created. Then, based on the created learning model and the measured vibration data, moving objects are detected. Specifically, based on the created learning model and the vibration data, the number of moving objects is detected.

[0046] As described above, in this embodiment, the moving object detection system creates a learning model based on the learning data by sliding the vibration data along the time axis by a predetermined time and extracting a predetermined length of data from the sliding time. This makes it possible to improve the detection accuracy of the moving object without performing long-term measurements or incurring costs.

[0047] Although an embodiment of the present invention has been described above, the forms to which the present invention can be applied are not limited to the above-described embodiment, and as exemplified below, appropriate modifications can be made without departing from the spirit of the present invention.

[0048] In the above-described embodiment, the moving body is a vehicle. However, the moving body may be other than a vehicle, or may include a vehicle. For example, the moving body may be a four-wheeled vehicle, a two-wheeled vehicle, or a pedestrian, and may be classified into a large four-wheeled vehicle, a medium four-wheeled vehicle, a small four-wheeled vehicle, a large two-wheeled vehicle, a medium two-wheeled vehicle, a small two-wheeled vehicle, or a pedestrian. The moving body may also be only a pedestrian.

[0049] In the above embodiment, the control unit creates multiple learning data by sliding 0.5 seconds along the time axis. This is not limiting, and the predetermined time for sliding may be adjusted so that the learning data overlaps with other learning data. For example, if the learning data has a predetermined length of 30 seconds, the time for sliding along the time axis may be changed as appropriate as long as it is within 30 seconds. In addition, the control unit does not need to create learning data by sliding, and may create learning data so that it overlaps with other learning data in terms of time. For example, the extraction start time of learning data 1 is set to 0 seconds as the measurement start time of learning measurement data, and 30 seconds of learning data is created. In addition, the extraction start time of learning data 2 is set to 0.5 seconds as the measurement start time of learning measurement data, and 30 seconds of learning data is created. In this way, the extraction start time and the predetermined length may be set in advance.

[0050] In the above embodiment, the control unit ends the process when it determines that the extraction end time of the extracted learning data matches the measurement end time of the learning measurement data. However, the process may also be ended when the extraction end time of the extracted learning data exceeds the end time of the learning measurement data. For example, the process may be ended when the extraction start time of the extracted learning data matches the end time of the learning measurement data. In this case, the control unit creates the learning data by treating the portion that exceeds the end time of the learning measurement data as silent data. [Industrial Applicability]

[0051] The present invention can be suitably adopted in a moving body detection system and a moving body detection method for detecting a moving body.

Claims

1. creating two or more pieces of learning data for creating a learning model from the learning measurement data, which is measurement data of a moving object, the learning data being partially overlapped in time with other learning data; creating the learning model for detecting a moving object based on the learning data; A moving object detection system characterized by detecting a moving object based on the created learning model and measurement data of the moving object.

2. The moving object detection system according to claim 1, characterized in that the learning data is created by sliding the learning measurement data along the time axis by a predetermined amount of time and extracting data of a predetermined length from the slid time.

3. The moving object detection system according to claim 1, characterized in that the learning data is created by extracting a predetermined length of data from the learning measurement data, some of which overlaps in time with other learning data.

4. The moving object detection system according to claim 1 , wherein the learning measurement data and the measurement data are vibration data.

5. measuring vibrations generated in a moving object when the moving body moves; 2. The moving body detection system according to claim 1, wherein measured vibration data is used as the learning measurement data.

6. measuring vibrations generated in a moving object when the moving body moves; Vibration peaks are extracted from the measured vibration data.

2. The moving body detection system according to claim 1, wherein the extracted vibration peaks are used as the learning measurement data.

7. The moving object detection system according to claim 1 , wherein the number of moving objects is detected based on the created learning model and measurement data.

8. The moving object detection system according to claim 1 , wherein the type of the moving object is detected based on the created learning model and measurement data.

9. The moving object detection system according to claim 1 , wherein the moving object is a pedestrian.

10. The moving object detection system according to claim 1 , wherein the moving object is a vehicle.

11. A sensor, A control unit, The control unit is creating two or more pieces of learning data from the learning measurement data, which is measurement data of a moving object measured by the sensor, for creating a learning model in which a part of the learning data overlaps with other learning data in time; creating the learning model for detecting a moving object based on the learning data; A moving object detection system characterized by detecting a moving object based on the created learning model and measurement data obtained by measuring the moving object.

12. creating two or more pieces of learning data from the learning measurement data, which are measurement data of a moving object, for creating a learning model in which a part of the learning data overlaps in time with other learning data; creating the learning model for detecting a moving object based on the learning data; A moving object detection method, comprising the steps of: detecting a moving object based on the created learning model and measurement data obtained by measuring the moving object.

Images