joshruby/nightjar-flight-20260804-night

nightjar — 2026-08-04 dusk/night sortie (lit drone)

The most informative flight of a five-sortie day. Same rig and site as the daylight sessions, flown at dusk with the drone's anti-collision light on, and — critically — with the detector switched from the daytime path to the night path.

The result

Detection in the 50–100 m band, which had refused to move all day:

sortie	conditions	detection @ 50–100 m
daylight #2	day detector, dark speck	0.5%
daylight #3	day detector, dark speck	0.0%
this one	night detector, lit point	10.0%

Tracking improved with it: TRACK 20.6% + COAST 22.1%, against 14.6% TRACK on the best daylight sortie. At 100–200 m detection is still 0%, so the light was not carrying that far.

The mechanism is the detector swap. By day a drone is a dark, low-contrast speck the daylight path must dig out of sky texture via ego-motion-compensated frame differencing — at 50–100 m it subtends ~2.9 px here. At night a lit drone is a bright point, which NightLightDetector finds by MAD bright-blob. Night mode also lifts the day-only cold-acquire restriction (day_yolo_live is False), so classical detections seed tracks natively.

⚠ detector.mode had been PINNED to day in the rig config since 2026-07-27 (auto flipped to night 965× on dark foliage at noon, since it thresholds on frame-median luminance and a canopy-filled frame is dark even at midday). It was set explicitly to night for this flight. Anyone reproducing this must set it — leaving it pinned means flying the whole night with the dark-speck detector and never using the light.

Envelope and caveats

Flight record in dji/. As with every 08-04 sortie after the first, this was a low flight close to the treeline, so the drone sat against foliage rather than clean sky and near the null of a horizontal array's elevation response (sensitivity goes as sin(el)).

Zoom stayed low: TRACK median 1.03×, p90 2.23×, max 3.91×, with 61% of tracked time still at full wide. That is containment behaving correctly rather than a fault — the tracked median range was 11 m, where the drone is already ~38 px and there is nothing to zoom for, and where angular rate (v/r) is far above the quiet-track gate the fast zoom law needs. The fast law still has not been exercised at the ranges it was built for.

A blinking-strobe hypothesis was tested and rejected. The acquisition gate needs 3 linked detections inside 1.5 s and drops a tracklet after a 0.7 s gap, so a ~1 Hz strobe looked like a plausible cause of poor tracking. The timing says otherwise: 842 detection frames (33.3%), gaps between detections median 0.34 s (p10 0.34, p90 0.35 — i.e. the blackbox sample period, so detections were essentially continuous), and 5 detections fit in the gate's 1.5 s window against the 3 required. The strobe did not starve acquisition.

Unresolved: it is not established that those detections were the drone rather than a ground light. The companion drone-free session nightjar-night-ambient-20260804, recorded the same evening at the same pointing, is the control for exactly that — anything the detector fires on there is by definition a false alarm.

Contents / gotchas

audio/uma16_<epoch>.wav   UMA-16, 16 ch, 48 kHz, S16_LE
audio/sb_<epoch>.wav      SB-POLARIS, 8 ch, 48 kHz, S32_LE
audio/cue_log.jsonl       cue gate log — thr_db + nj_state per candidate
video_segments.tar        1080p native + 768x432 substream, 60 s MPEG-TS
blackbox/state.jsonl      tracker FSM ~2.9 Hz    blackbox/frames.tar  substream stills
dji/*.csv, *.txt          decoded + original FlightRecord (CUSTOM.dateTime = true UTC)

UMA-16 0-based channel 9 is a dead electronic-floor channel — exclude it, run on 15. SB-POLARIS has 3 live capsules of 8; detect them by RMS (live ~−70 dBFS, dead ~−190) rather than hardcoding — the live set has changed between sessions. ⚠ The two arrays use different sample widths (UMA S16_LE, SB S32_LE); reading both as int16 scrambles SB channel identity and destroys its spectrum.

Calibration as flown: az_offset_deg 301.37, invert_az true (re-solved from the JBL beacon that afternoon, 0.40° drift). detector.mode: night. Cue producer: cheap daemon (SUBHUNT=0), site notch 100–200 Hz, state-dependent presence gate (14.6 dB while TRACK, 10.0 dB while searching). Start pose pan 178.9 / tilt 45. Cue-gated blob seeding present in the build but disabled (and moot at night, where classical detections already seed).

GPS-referenced pixel work on this rig floors out at about 5° of bearing error — the DJI's own ±2–5 m is 4–9° of bearing at the 20–45 m ranges where optical correspondences live. Treat in-frame or pixel-level claims derived from GPS as approximate.