1:45 PM - 3:15 PM
[O08-P88] A real-time detection system for meteors using machine learning with video and radio waves
Keywords:Meteor, Machine learning
1. Background and purpose of the study
Our school's Astronomy and Meteorology Club has been conducting visual observations of the Perseid meteor shower since 1953, and is currently observing the Geminids meteor shower as well. However, since it is burdensome to increase the number of all-night observations, we decided to construct an automatic observation system using video and radio waves. In both cases, we also worked on machine learning to identify meteors using YOLO, and developed a system that notifies smartphones and other devices in real time.
2. Methods
The astronomy club developed an automatic meteor observation system using two methods video observation and radio wave observation. In addition, a program was developed to detect meteors in real-time using machine learning.
3. Construction of a video observation device and development of a detection program
Five security cameras were installed facing four directions and the zenith direction, and a program was developed to synthesize them in a way that covers the entire sky.
The detection program was developed, which uses machine learning to detect meteors in YOLOv9. A dataset was created from 200 images containing meteors and performed machine learning. Then, to avoid false detection of satellites, those images were excluded if they were detected in more than 3 images per minute. As a result, the meteor detection was successful, and both the fit rate and the reproduction rate were improved. However, this program sometimes misses meteors when there are multiple meteors in one minute. Therefore, a method to determine the presence or absence of satellites by comparing the IoU of the 1-minute and 5-second composite images is currently being developed.
4. Development of radio wave observation equipment and detection programs
The radio wave observation uses the 53.755 MHz radio wave transmitted for meteor observation. The astronomy club developed a program to automatically detect meteors from echo images obtained from the free software MROFFT. A model was developed to discriminate between meteors using YOLOv8, which performs object detection. The target frequency is cut out of the image, the meteor is discriminated against by object detection, and the time when the meteor occurred is obtained. The model was actual created and successfully discriminated.
5. Development of a real-time detection system
The astronomy club developed a program to execute the detection programs described in 4 and 5 in real-time, and when a meteor is detected, the time of observation and the image of the detected meteor are stored in the NAS. In addition, as shown in Figure 1, the video observation succeeded in notifying the appearance of meteors to the communication application Discord in real-time.
6. Results and discussion of the Orionids meteor shower in 2023 observed by video, radio wave, and visual observation
The Orionids meteor shower of 2023 was compared by video, radio waves, and visual observation. The video and radio observations were the results of visual confirmation. Figure 2 shows that video observations became extremely large after the sky cleared. The reason for the increase in the number of video observations is thought to be that the infrared security camera captured meteors of low magnitude, which are difficult to confirm visually. Also, the number of detections by radio observation was lower than expected. This is thought to be due to the fact that many meteors in the Orionids meteor shower have high ground speeds.
7. Summary and Future Prospects
The astronomy club built video and radio observation devices and developed a meteor detection program using machine learning. The observation results showed that video observations captured more meteors than radio and visual observations. In the future, the astronomy club aims to complete a real-time meteor detection system for radio wave observations, and will also improve the accuracy of radio observations.
Our school's Astronomy and Meteorology Club has been conducting visual observations of the Perseid meteor shower since 1953, and is currently observing the Geminids meteor shower as well. However, since it is burdensome to increase the number of all-night observations, we decided to construct an automatic observation system using video and radio waves. In both cases, we also worked on machine learning to identify meteors using YOLO, and developed a system that notifies smartphones and other devices in real time.
2. Methods
The astronomy club developed an automatic meteor observation system using two methods video observation and radio wave observation. In addition, a program was developed to detect meteors in real-time using machine learning.
3. Construction of a video observation device and development of a detection program
Five security cameras were installed facing four directions and the zenith direction, and a program was developed to synthesize them in a way that covers the entire sky.
The detection program was developed, which uses machine learning to detect meteors in YOLOv9. A dataset was created from 200 images containing meteors and performed machine learning. Then, to avoid false detection of satellites, those images were excluded if they were detected in more than 3 images per minute. As a result, the meteor detection was successful, and both the fit rate and the reproduction rate were improved. However, this program sometimes misses meteors when there are multiple meteors in one minute. Therefore, a method to determine the presence or absence of satellites by comparing the IoU of the 1-minute and 5-second composite images is currently being developed.
4. Development of radio wave observation equipment and detection programs
The radio wave observation uses the 53.755 MHz radio wave transmitted for meteor observation. The astronomy club developed a program to automatically detect meteors from echo images obtained from the free software MROFFT. A model was developed to discriminate between meteors using YOLOv8, which performs object detection. The target frequency is cut out of the image, the meteor is discriminated against by object detection, and the time when the meteor occurred is obtained. The model was actual created and successfully discriminated.
5. Development of a real-time detection system
The astronomy club developed a program to execute the detection programs described in 4 and 5 in real-time, and when a meteor is detected, the time of observation and the image of the detected meteor are stored in the NAS. In addition, as shown in Figure 1, the video observation succeeded in notifying the appearance of meteors to the communication application Discord in real-time.
6. Results and discussion of the Orionids meteor shower in 2023 observed by video, radio wave, and visual observation
The Orionids meteor shower of 2023 was compared by video, radio waves, and visual observation. The video and radio observations were the results of visual confirmation. Figure 2 shows that video observations became extremely large after the sky cleared. The reason for the increase in the number of video observations is thought to be that the infrared security camera captured meteors of low magnitude, which are difficult to confirm visually. Also, the number of detections by radio observation was lower than expected. This is thought to be due to the fact that many meteors in the Orionids meteor shower have high ground speeds.
7. Summary and Future Prospects
The astronomy club built video and radio observation devices and developed a meteor detection program using machine learning. The observation results showed that video observations captured more meteors than radio and visual observations. In the future, the astronomy club aims to complete a real-time meteor detection system for radio wave observations, and will also improve the accuracy of radio observations.