۱- Introduction

۲- Image processing pipeline

۳- Fast meteor detection

۴- Classification via deep learning

۵- CNN as a meteor detector

۶- Leading edge position refinement

۷- MULTI-SITE aggregation of tracks

۸- Atmospheric trajectory estimation

References

Abstract

The processing pipeline for both video meteor detection and track analysis has evolved to embrace several new algorithms, which have improved the efficiency and performance of various steps in the meteor image processing chain. With the advent of larger pixel count digital sensors, the image processing techniques have needed to keep up with the computational load by not only employing higher end processors, but developing faster thresholding, clustering, and tracking algorithms for detection. In addition, machine learning methods employing both recurrent and convolutional deep neural networks have helped remove the human-in-loop false alarm mitigation step inherent in many meteor collection processing streams. The application of a matched filtering algorithm has helped to refine the measurement positional accuracy of propagating meteor tracks for post-detection analysis. The use of improved multi-site track aggregation has dramatically reduced the occurrence of mis-associating unrelated tracks during the combination into a single trajectory. When coupled with an improved minimization metric in the multi-parameter fitting method for trajectory estimation, this yields better meteor orbital solutions. Finally, proposed concepts in using a convolutional neural network as a meteor detector and performing trajectory fitting with an empirically based propagation model, show promise for more robust meteor image processing and analysis in the near future.

Introduction

The meteor imaging community has begun migrating towards multimegapixel, progressive-scan digital sensors, and away from the traditional analog cameras with less than half-million pixels per frame. While this has been beneficial from the lowered noise, image quality, higher bit depth, spatial coverage, and angular resolution perspectives, the significantly larger image sizes have stressed processing loads. For example, full high-definition (HD) 1080p imagery possesses what amounts to approximately six times the pixel count of NTSC or PAL video. The computational issue has been partially offset by both Moore’s law as well as employing larger capacity CPUs, but this can be limiting in multicamera systems without resorting to high cost PC systems or custom GPU implementations. In addition, the advent of innovative meteor collection systems for tracking meteors for fragmentation studies and spectroscopy that require real-time responsiveness of these short-lived events, has resulted in a re-examination of the entire image processing chain to increase computational efficiency, while also maintaining detection and analysis robustness. Finally, with the exploding growth of low-cost meteor camera deployments, meteor collection networks have tried to automate as many steps in the processing pipeline as possible, specifically those steps which were traditionally addressed by having a human-in-the-loop (HIL) perform the process. Thus, several algorithms have been undergoing new development and are being incorporated into various meteor imaging systems and their processing pipelines. This paper addresses a broad base of those enhancements and will highlight the significant improvements in algorithms, providing references for those algorithms described in greater detail in the published literature.