Automated detection of meteor trails in all sky images using Radon transformation and convolutional neural network

We present an automated pipeline for detecting meteor trails in single-frame, half-minute-cadence all-sky images acquired by the Digital Autonomous Fireball Observatories (DAFO) network. Unlike existing approaches that rely either on handcrafted feature extraction or purely data-driven methods, the proposed system combines classical image processing techniques (Radon transformation and Fourier-transformation-based filtering) with a convolutional neural network (CNN) for robust classification of candidate trails. The method is specifically designed for fish-eye camera images and operates without dependence on a particular hardware configuration. It addresses limitations of our previous approach Suk and Šimberová [ 1 ], namely high false-positive rates and insufficient robustness to ambiguous trails such as blinking satellites and airplanes. The proposed pipeline was evaluated on data from eleven nights across multiple observatories. For clearly identifiable meteor events, the method achieves a precision of 93.9% and a recall of 95.3%, with near-complete detection of all observable meteors not obscured by ground objects. The system also significantly reduces the number of false detections compared to earlier methods. These results demonstrate that combination of classical detection techniques with modern neural network classification provides an efficient and computationally feasible solution for large-scale automated meteor monitoring.