|FMDT| is derived from a software which was designed to detect meteors on board |ISS| or a |CubeSat|. |FMDT| is foreseen to be applied to airborne camera systems, e.g. in atmospheric balloons or aircraft. It is robust to camera movements by a motion compensation algorithm.
FMDT is ready for real-time processing on small boards like Raspberry Pi 4 or Nvidia Jetson Nano for embedded systems. For instance, on the Raspberry Pi 4 (@ 1.5 GHz), |FMDT| is able to compute 30 frames per second on a |HD| video sequence while the instant power is only around 4 Watts.
Exemple of meteors detection and visualization.
:numref:`fig_detection_chain` shows an example of detection on one frame. Green |BBs| represent detected meteors, purple |BBs| represent detected stars and orange |BBs| represent detected noise (= something which is not a meteor and not a star).
The detection chain.
:numref:`fig_detection_chain` presents the whole |FMDT|'s detection chain. For each pair of images, an intensity hysteresis threshold, a connected component labeling and an analysis algorithm are applied to get a list of |CCs| with their bounding boxes and surface. Moreover, it also provides the first raw moments to compute the centroid (x_G,y_G)=(S_x/S,S_y/S) of each blob of pixels. A morphological threshold is then done on the surface S to reject small and big |CCs|. A |k-NN| matching is applied to extract pairs of |CCs| from image I_{t+0} and I_{t+1} with t the image number in the video sequence. These matches are used to perform a first global motion estimation (rigid registration). Note that |CCs| are sometimes refered as |RoIs| in this documentation.
This motion estimation is used to classify the |CCs| into two classes - still stars or moving meteors according to the following criterion: |e_k-\bar{e_t}| > \sigma_t with e_k the compensation error of the |CC| number k, \bar{e_t} the average error of compensation of all |CCs| of image I_t and \sigma_t the standard deviation of the error. A second motion estimation is done with only star |CCs|, to get a more accurate motion estimation and a more robust classification. Finally a piece-wise tracking is done by extending the (t+0,t+1) matching with (t+1,t+2) matching (and so on) to reduce the amount of false positive detection.
|IMCCE| astronomers (from Paris's Observatory) led an airborne observation campaign of the 2022 \tau-Herculids. The 2022 \tau-Herculids mission is detailed here. The data collected by the mission have been processed with |FMDT|. The detection results helped the astronomers to see more meteors than their first "manual" detection (by human eyes). From 28 to 34 meteors thanks to |FMDT| automated detection. Detailed results are available in an article published in the Astronomy & Astrophysics journal :cite:`Vaubaillon2023`.
Some preliminary results about the parallel implementation of the detection chain (see :numref:`fig_detection_chain`) have been presented in a poster :cite:`Kandeepan2022` of the workshop |AFF3CT|. The poster shows results in terms of throughput (|FPS|), latency and energy consumption. The selected hardware targets match embedded systems constraints (e.g. \mathcal{T} \ge 30 |FPS| and \mathcal{P} \leq 10 Watts).
.. bibliography:: ../refs.bib