Luis Varela, New Mexico State University; Laura Boucheron, New Mexico State University; Nick Malone, Tau Technologies; Nicholas Spurlock, Tau Technologies
Keywords: Streak detection, image processing, convolutional neural network (CNN), deep learning, Wide field of View, night sky imagery, meteor detection
Abstract:
In many applications it is desirable to detect and localize moving objects (e.g., satellites or meteorites) in images of the night sky. These applications range from mission critical tasks such as space situational awareness and satellite catalog maintenance to astrophysics and meteor impact estimation. Low signal-to-noise ratio (SNR) due to faint objects, light pollution, or atmospheric conditions complicates automated algorithms for such analysis. Since a moving object will trace out a streak in long-exposure images, we present and study a new streak detection method based on deep learning, specifically the You Only Look Once (YOLO) Convolutional Neural Network (CNN), to provide more robust performance across varying image quality. This novel streak detection algorithm is tested on a multi-camera wide field of view system (WFoV). This system has a 60 – 160 degree field of view and is a fixed, staring persistent array. It can reach magnitude 13-20 objects and captures 10 second frames all night, producing 4 Terabytes of data per night. We compare our results to two existing methods based on the Hough Transform, and the Phase Congruency Transform, each of which demonstrates different regimes of performance depending on imaging conditions. It is possible that the learning capabilities of the CNN will result in a method that can span these different performance regimes. In this paper, we compare accuracies in streak detection between the two baseline methods and our CNN streak detector. Since human labeling of images is time consuming, we used a data curation to clean a dataset for training and validation. We additionally study cases where the baseline methods fail and analyze the constraints of each method.
Date of Conference: September 17-20, 2019
Track: Machine Learning for SSA Applications