VigiSky Sentinel

Sentinel downloads the Celestrak catalog (visual + stations), caches it, and computes passes with SGP4 on your PC. Morning review and the authenticity stamp work offline once a TLE set has been fetched. ↻ refreshes the catalog; you can also import a .tle file.

The old naive detector (threshold on the frame mean) is gone. Photos, demo skies, Cosmos hunt, camera watch and stacked video all go through the same astronomical pipeline: local background, matched filter, classified sources. One method, one verdict style.

CUDA stays the fast path on NVIDIA. Without a GPU, or compiled without the toolkit, Sentinel runs multi-thread on the CPU: same image analysis, same video tracking. The galactic Lab still needs the CUDA build.

Load your sighting video: every moving object becomes a dated event with its trajectory, angular speed and light curve. Aircraft strobe lights, satellite steadiness, meteor brevity — the engine recognizes them, and says so cautiously. It also measures authenticity clues (photon noise, micro-shake, frame cadence) to flag AI-generated or heavily processed videos — without ever crying fake. → The method explained

Sentinel recognizes the stars in your photo and derives exactly where the lens was pointing: field center to the arcminute, scale, rotation — computed locally, using the same mathematics as satellite star trackers (triangle matching). Stars get named on the image, a celestial grid overlays it, and any point in the photo receives measured coordinates. An unexplained light then has a position and angular size proven by the stars around it. The night watch anchors itself: the camera knows where it points, every event gets a real azimuth/elevation, and the satellite counter-check verifies the time and the direction. The measurement residual is always displayed — a weak solve says "uncertain", never the opposite.

One button in the header switches the whole application to English: interface, night watch, morning review, counter-checks. Your preference is remembered between launches. French remains the reference language — English is a display layer; case files and data are unchanged.

The star solver gains a rescue pass: when the standard sweep fails (badly wrong field hint, star-poor sky), it retries with extended bins then in depth — more stars, more triangles — within the same time budget. On the bench: 98% of fields solved "solid" under 30% impostors, 20% missing stars and 2 px of noise, zero false locks. Fields that already solved pay nothing.

Plug in an all-sky fisheye camera and the night watch solves the complete celestial dome: the solver automatically switches to the equidistant projection (r = f·θ), self-calibrates the focal length on the stars, and every event gets its measured azimuth/elevation — bench-validated to 0.18°. Combined with the marathon, that's a station watching the whole sky, every night, on its own.

Tick "Marathon" and leave for the weekend: at dawn, Sentinel closes the night and writes the morning review to disk; at dusk, the watch restarts by itself (local solar computation, validated to the degree). Power cut, camera cable unplugged? The station re-arms and resumes its night where it left off. A multi-night recap accumulates surveillance hours and recurrences — a true autonomous observation station.

Your stacked deep image now exports as 32-bit FITS, astronomy's scientific format — and when the field is star-anchored, the astrometric WCS is embedded in the header: the file opens already plate-solved in Siril, DS9 or PixInsight, every pixel knowing its celestial coordinates. Sentinel now speaks the language of observatories.

After a successful star anchoring, one click saves your lens's EXACT measured field as a named profile ("Garden dome", "S23 wide"…). From then on, everywhere — photo analysis, video, night watch — pick the profile instead of guessing: the solver starts from the true field, faster and safer. The foundation of multi-camera support.

Pick a folder: Sentinel analyzes every photo sequentially (up to 500), sorts the summary by interest — candidates first — and exports everything as CSV for your own tools. An observing station's routine, without touching each image.

One click on any tracked object and Sentinel generates a professional, print-to-PDF dossier: every measurement with its error margins (star-anchored rate, orbit test, light fingerprint, authenticity verdicts), the original file's SHA-256 fingerprint for evidence integrity, and a full methodology annex with the validation-bench results. The document you hand to an investigator, a journalist, a forum — generated locally, nothing leaves your PC.

The app's last honest refusal just fell: all-sky fisheye lenses (real watch-station hardware, 120-185° fields) are now star-solved — equidistant projection, focal length auto-calibrated to 0.2%, centers within arcminutes. And if your lens sits in the ambiguous zone (~100-140°), the app tries both models before giving up. A dome camera becomes a Sentinel station.

When two witnesses triangulate a fireball at the end of its luminous flight, Sentinel integrates the dark flight: gravity, atmospheric drag and winds, fragment by fragment. The 10-gram pieces drift with the wind, the 10-kilo ones carry far: the fall zone takes shape, mass by mass, with coordinates. The day a fireball crosses the sky, the network's videos will be able to say where to look for the stone.

Sentinel now archives your watches and cross-references the nights. An object seen several nights in the same region of the sky, with a steady time shift (−4 min/day…): that is the orbital signature — a satellite, even an uncataloged one, is far more likely than a one-off phenomenon. Exactly fixed schedule? Probably a scheduled flight. The morning review tells you on its own: "seen 5 nights around 23:40, shifting −3.9 min/day, predictable to ±1 min".

When you wake up, one button in the morning review: "Publish my night". Your watch joins the network's public page: satellites, aircraft, meteors and signals to re-examine, counted and cross-checked, station by station. Voluntary and anonymous: position rounded to ~11 km, never an address, no images. The sky, watched every night, told every morning.

Every moving object now gets its fingerprint: the autocorrelation of its light curve extracts the blinking period. A tumbling rocket body beats at a steady 3 seconds, aircraft strobes at 1 Hz, a cruising Starlink is flat, a lantern flickers with no rhythm. "Periodic blinking: 3.00 s (12 periods seen)" — one more physical clue in the case file, never a magic identification.

"What if the object emitted no light at all?" Here is the answer. During every video analysis, Sentinel monitors the flux of the field's fixed stars: a dark object crossing extinguishes them one by one — a sequence of extinctions aligned along a constant-velocity front. A cloud also blots out stars, but wide, diffuse and disorderly: it is recognized and rejected, as is scintillation. Even a black object leaves a trace — the stars it hides.

A lone witness cannot know the distance of a light in the sky. Two witnesses aiming at it at the same moment can: their sightlines cross in space. Sentinel now triangulates the network's full trajectories: real altitude with error bounds, speed in km/h, and the parallax verdict — "object too distant for this baseline, spread the stations out" when the geometry cannot conclude. Witness positions never leave the server: the geometry travels in a relative frame.

A star-anchored video trajectory gives no distance… unless the object follows an orbit. Sentinel rebuilds the 3D track for every possible altitude: at the right one — and only there — the measured angular rate matches the Keplerian rate and the motion is planar and uniform. Result: "compatible circular orbit: ~420 km, period 92.8 min". An aircraft, too irregular, fits no altitude — and the app says so. Textbook celestial mechanics, on a phone and a PC.

A starry-sky video claims it was filmed somewhere, at some time? Sentinel recomputes the sky of that very moment and confronts it: was the identified star field above the horizon? Is the sky's rotation (0.25°/minute) measurable in the video — composites never have it? Was it daylight? Should Jupiter or the Moon have been in frame? Each test returns its own verdict — compatible, contradicted, or untestable — never a magic "real/fake". The method explained: proof by the stars. v2: expected bodies are now verified in the frame (is a bright source measured at the predicted pixel? Jupiter missing = contradiction), and the video's moving object is cross-checked against cataloged satellites at the claimed time — per-test verdicts, as always.

Real lenses bend the sky: barrel or pincushion distortion, 1–4 % at the edge of a phone's field. While anchoring to the stars, Sentinel now measures that distortion on your own photo — and corrects it. The panel says it plainly: "barrel distortion 2.1% at the corner — measured and corrected". On a clean lens nothing changes: the correction only engages when it clearly improves the residual; otherwise the app tells you "no significant lens distortion detected". v2.1: edge-of-field stars pushed away by distortion are re-captured once the model is measured — more matched stars, better accuracy exactly where the lens bends most.

The morning review already cross-checked every event against satellites — it now also checks real air traffic (sampled during the night via OpenSky Network): "✈️ AFR123 was in that direction, altitude 11,000 m". And for brief streaks, the active meteor showers: "☄️ consistent with the Perseids — radiant 40° up, near peak". Satellites, aircraft, meteors: the three usual suspects, checked automatically every night against real data.

Film the sky for 20 seconds: Sentinel aligns every frame on the stars (to a tenth of a pixel) and stacks them. Noise collapses with the square root of the frame count — stars invisible in one frame appear across three hundred. Two outputs: the deep image (aligned mean, analyzable in the studio and star-anchorable) and the trails (per-pixel maximum — moving objects draw their full path). The astronomers' "lucky imaging" technique, in one click.

Analyze a video then anchor it to the stars: every moving object gets a trajectory in true celestial coordinates and its measured angular velocity (no more assumed-field estimates). One click exports the trajectory to the VigiNet network in the witness-sighting format: it becomes identifiable and triangulatable as if a witness had tracked it by phone — at instrument grade. Tested on a real ISS pass geometry: direct match at 0.4°.

Like sailors with a sextant, but automatic: lay the phone flat facing the sky, take 2–5 photos, and Sentinel recognizes the stars at your zenith to derive your latitude and longitude — no GPS, no network. The point of sky directly above you has a declination equal to your latitude and a right ascension equal to your local sidereal time: a three-century-old formula, solved by the stars. Honest accuracy shown (it depends mostly on how level the device is): an educational sextant and an offline fallback.

Plug in a camera — a webcam or an all-sky fisheye up to 185° — and Sentinel watches the sky all night. In the morning: the morning review — everything that moved, classified, with snapshots. Give the station's position and every event is automatically cross-checked against the night's bright satellite passes (Celestrak data): "STARLINK was passing at that time, culminating at 45°" — the book's chapter 7 counter-investigation, automated. A time match is a clue, never a confirmation.

The Deep Sky Engine analyzes your photo like an observatory: local sky background, extraction of every light source (sub-pixel centroid, FWHM, elongation), cautious classification — star, satellite trail, hot pixel, or signal worth a second look. A 24 Mpx photo takes ~0.2 s on an NVIDIA GPU; without one, the CPU takes over.

Each observation becomes a numbered, exportable investigation sheet: witness data, checks performed, analysis, verdict — and its limits, stated in plain words. Never "UFO detected".

A CUDA galaxy simulator — one million stars, dark matter, the Milky Way-Andromeda collision, ring galaxies, live supernovae — to understand what "astronomical unknown" really means.