Skip to main content

Glossary

A working reference for satellites, measurements, formats, and the terms specific to this repo.

1. Satellites and signals

  • GNSS — Global Navigation Satellite System. Umbrella term for all constellations.

  • Constellation — one operator's satellite fleet:

    • GPS (USA, sat ids G01…)
    • GLONASS (Russia, R)
    • Galileo (EU, E)
    • BeiDou (China, C)
    • QZSS (Japan, J)
    • NavIC (India, I)
    • SBAS (augmentation, S)

    The ZIM3 file has 55 sats because it tracks all of them.

  • PRN / sat id — satellite number inside a constellation. G04 = GPS satellite 4.

  • L1 / L2 / L5 / E5 / B2 — radio frequencies (bands). L1 ≈ 1575.42 MHz, L2 ≈ 1227.60 MHz, L5 ≈ 1176.45 MHz. Multi-frequency helps because ionosphere delay depends on frequency, so two bands let you cancel it.

  • Signal — a band plus a modulation code, e.g. GPS_L1CA = GPS L1 Coarse/Acquisition. One satellite emits several signals.

2. What the receiver actually measures

  • Observation (obs) — one measurement of one signal from one satellite at one time.
  • Epoch — one time instant containing all observations. These files have one epoch every 30 s (ZIM) or 15 s (ab02).
  • Pseudorange — travel-time distance satellite-to-receiver, in metres. Called pseudo because receiver clock error is baked in, so it's off by a common bias, typically hundreds of km. Noise ~0.3–3 m. Fields C1C, P2, C1 in RINEX.
  • Carrier phase — distance measured by counting carrier wave cycles. Noise ~2 mm, 1000× better than pseudorange. Problem: you only know the fractional part; the integer number of whole cycles is unknown. That unknown is the ambiguity. Fields L1, L2.
  • Doppler — frequency shift, gives velocity directly.
  • SNR / CN0 — signal strength in dB-Hz. ~45 = strong, <30 = weak or reflected. Fields S1, S2.
  • Cycle slip — receiver loses phase lock for an instant, integer ambiguity jumps. Must be detected or position breaks.
  • LOS / NLOS — Line Of Sight / Non-Line Of Sight. LOS = direct path sat-to-antenna. NLOS = direct path blocked, only a reflected signal arrives, which is longer, so range is biased metres to tens of metres too long. Main urban error source, and exactly what the robot-body mask kills.
  • Multipath — direct signal plus reflection arrive together, interfering. Related to NLOS but not identical.

3. Error sources

  • Ionosphere — charged upper atmosphere, delays code / advances phase, 1–30 m, frequency-dependent.
  • Troposphere — lower wet/dry atmosphere, ~2.3 m at zenith, grows as 1/sin(elevation). Not frequency-dependent, so must be modelled (Saastamoinen) or estimated.
  • IFLC / ionosphere-free combination — arithmetic mix of two frequencies that cancels ionosphere. Costs noise.
  • Ephemeris — the orbit model of a satellite. Needed to know where the sat was when it transmitted.
  • Broadcast ephemeris — orbit data transmitted by the satellites themselves. Accuracy ~1–2 m. This is the brdc / .nav / MN.rnx file.
  • Precise products — computed after the fact by analysis centres:
    • SP3 = precise orbits (~2.5 cm)
    • CLK = precise satellite clocks
    • IONEX = global ionosphere maps
    • DCB / OSB / Bias-SINEX = hardware code biases per satellite/signal
  • SSR — State Space Representation. Orbit/clock/bias corrections streamed in real time instead of computed later. CLAS and MADOCA are Japan's QZSS SSR services, broadcast on the L6 signal.

4. File formats

RINEX — Receiver INdependent EXchange format. Plain text, the universal GNSS interchange format. Two kinds matter here:

  • observation file — the measurements
  • navigation file — the ephemerides

RINEX 2 vs RINEX 3 — version 2 is old, cramped layout, one global list of observation types. Version 3 is modern, per-constellation observation types, better multi-GNSS. The parser bug hit above is in the version 2 path only.

File naming, old (RINEX 2) — ssssDDD0.YYt

Example Meaning
ab022370.26o station ab02, day-of-year 237, session 0, year 26, type o = observation
brdc2370.26n broadcast, DOY 237, year 26, n = GPS navigation
.26g / .26p GLONASS nav / mixed nav

File naming, new (RINEX 3) — long form

ZIM300CHE_R_20262370000_01D_30S_MO.rnx

Field Meaning
ZIM3 station
00 monument/receiver number
CHE country (Switzerland)
R from data centre archive
2026 237 0000 start year, DOY, hour+minute
01D 1 day span
30S 30 second sampling
MO Mixed constellation Observation (MN = Mixed Navigation)
.rnx plain RINEX text

Other formats

  • Hatanaka / CRINEX / .crx — RINEX observation compressed by a GNSS-specific algorithm (stores differences between epochs). Halves size before gzip. Must decompress with crx2rnx or Python hatanaka before this library reads it. .crx.gz = Hatanaka then gzip.
  • .pos — solution output. One row per epoch: week, tow, ECEF XYZ, lat/lon/height, quality flag, satellite count, DOP, extra diagnostics. Convention from RTKLIB.
  • Also in this lib — RTCM (binary correction/observation streaming standard, used by NTRIP casters), UBX (u-blox receiver binary), SBF (Septentrio binary), NMEA (old ASCII sentences, $GPGGA…), BINEX (binary exchange), KML (Google Earth track).

5. Time

  • GPS week — weeks since 1980-01-06. This data: 2433.
  • tow — Time Of Week, seconds since Sunday 00:00 GPS time, range 0…604800. 172800.000 = 48 h into the week = Tuesday 00:00. Week + tow uniquely stamps an epoch. It's the join key in the los_prob CSV because it's what the solver holds internally.
  • Leap seconds — GPS time does not have them, UTC does; currently 18 s apart.

6. Coordinate frames

  • ECEF — Earth-Centred Earth-Fixed. XYZ metres, origin at Earth's centre, rotates with Earth. What the solver works in. Zimmerwald ≈ (4331300, 567537, 4633133).
  • Geodetic / LLH — latitude, longitude, height above the reference ellipsoid WGS84. Same point, human-readable.
  • Ellipsoidal vs orthometric height — ellipsoid height is geometric; orthometric is height above the geoid (mean sea level). Differ by tens of metres.
  • ENU — East, North, Up. Local flat-earth frame tangent at your position. Good for "how far did I move" and for sky directions.
  • NED — North, East, Down. Same idea, aviation convention. Library exposes velocity in both.
  • Azimuth — compass direction to a satellite, degrees, 0 = North, increasing clockwise (90 = East).
  • Elevation — angle above local horizon, 0 = horizon, 90 = straight up (zenith).
  • Elevation mask — reject satellites below N degrees. Low satellites travel through more atmosphere and reflect off everything. Default 15° here.
  • Attitude — orientation of your robot: roll (tilt sideways about forward axis), pitch (nose up/down), yaw (heading). ZYX means apply yaw, then pitch, then roll.
  • REP-103 / FLU — ROS convention: body x Forward, y Left, z Up; world frame ENU. What body_mask.py uses.
  • Quaternion — 4-number rotation representation, no gimbal lock. include/libgnss++/fusion/attitude.hpp uses these.

7. Positioning methods

Method Description Accuracy Binary
SPP Single Point Positioning. One receiver, pseudoranges only, broadcast ephemeris. Fast, always works, no base station. 1.5–10 m gnss_spp
DGNSS Differential; base station sends corrections, code only. ~0.5–2 m —
RTK Real Time Kinematic. Two receivers, carrier phase, resolve the integer ambiguities. Needs a base within roughly 10–40 km. 1–2 cm when fixed gnss_solve
PPP Precise Point Positioning. Single receiver, but precise orbit/clock products instead of a base. Takes 10–30 min to converge. Few cm gnss_ppp
PPP-RTK / CLAS / MADOCA PPP plus atmospheric corrections from a network, so it converges fast like RTK without a nearby base. — —

8. RTK vocabulary

  • Base / rover — base is the station on a known point, rover is the one you're positioning. In this test ZIM2 = base, ZIM3 = rover.
  • Baseline — vector (and distance) base-to-rover. Shorter = better; errors cancel because both antennas see the same atmosphere. The Zimmerwald pair is co-located, so essentially the easiest possible case, hence 96.67% fix.
  • Single difference (SD) — rover measurement minus base measurement to the same satellite. Cancels satellite clock and most atmosphere.
  • Double difference (DD) — difference of two single differences, using a reference satellite. Also cancels receiver clocks and leaves an integer-valued ambiguity. This is the core of RTK, and it's why the exclusion guard protects "a system's last reference satellite" — remove it and that whole constellation's DD set collapses.
  • Ambiguity resolution (AR) — determining the integer cycle counts. LAMBDA is the standard search algorithm.
  • Float solution — ambiguities left as real numbers. Decimetre level. Quality flag Q=2 in .pos (here 3 also appears for other modes).
  • Fixed solution — ambiguities resolved to integers. Centimetre level. Q=1 classic RTKLIB, Q=4 in this library's output.
  • Ratio test — best-vs-second-best candidate ratio; above threshold (default ~3) the fix is accepted.
  • Fix rate — fraction of epochs that reached a fixed solution. Baseline here: 96.67%.
  • Wrong fix — fixed integers that are wrong. Dangerous: confidently reports centimetres while being metres off. The README tables track it separately for that reason.
  • Fix-and-hold — once fixed, constrain the ambiguities to keep them.

9. Filters and estimation

  • Least squares / WLS — solve for position from redundant measurements; the weighted version trusts high-elevation, high-SNR measurements more.
  • Kalman filter (KF) — recursive estimator carrying position and ambiguities forward through time with a covariance.
  • FGO — Factor Graph Optimization. Instead of one forward pass, optimise a window of recent epochs jointly. Better in bad environments, more compute. GTSAM is the library used. gnss_fgo.
  • Tightly coupled GNSS/IMU — fuse raw satellite measurements with the IMU inside one estimator, rather than fusing final positions. Survives short outages.
  • RAIM / FDE — Receiver Autonomous Integrity Monitoring / Fault Detection and Exclusion. Leave one satellite out at a time, see if residuals improve; if so, that satellite is faulty. Automatic outlier removal.
  • Residual — measured minus predicted. Small residuals = consistent solution.
  • Robust / Huber weighting — downweight rather than delete large residuals.

10. Quality numbers

  • DOP — Dilution Of Precision. Pure geometry factor: how the satellite spread amplifies measurement error. Lower = better. GDOP geometric, PDOP position, HDOP horizontal, VDOP vertical. Under 2 = excellent, over 10 = poor. The 40° mask run hit GDOP 29 because only 4 satellites clustered overhead remained — exactly the failure mode aggressive robot masking can cause.
  • Availability — fraction of epochs with any valid solution.
  • RMS — root-mean-square error against truth.
  • P95 — 95th-percentile error. Says what your bad epochs look like, unlike RMS which averages them away.
  • pp — percentage points (difference of two percentages).

11. Networks, stations, archives

  • CORS — Continuously Operating Reference Stations, the US NOAA/NGS network. Free, open HTTP, no login. Source of ab02, p091.
  • IGS — International GNSS Service. Global scientific network, produces the precise SP3/CLK products.
  • EUREF — the European reference network.
  • BKG — German federal agency running an open IGS/EUREF data mirror. Where the Zimmerwald files came from.
  • Zimmerwald — an observatory near Bern, Switzerland. ZIM2 and ZIM3 are two receivers at that same site, a few metres apart, which is why RTK fixes almost every epoch. Good property for testing: any degradation you see is caused by your mask, not by the baseline.
  • brdc — "broadcast", the daily merged broadcast-ephemeris file.
  • NTRIP — protocol for streaming RTCM corrections over HTTP. Caster = the server. Live RTK path, gnss_stream / gnss_live.
  • DOY — day of year, 001–366. 237 = 2026-08-25.

12. Terms specific to this repo

  • los_prob — the number in the mask CSV. 1.0 = trust as line-of-sight, 0.0 = certain NLOS. Not from the library — you supply it. Modes: OFF ignore, TWO_TIER inflate sigma below a threshold, CONTINUOUS scale sigma smoothly, EXCLUDE hard-drop.
  • Sigma inflation — multiply a measurement's assumed noise so the filter distrusts it. Softer than exclusion; it can't create a geometry hole.
  • --nlos-min-sats — guard: refuse to apply exclusions if too few satellites would survive.
  • Selection snapshot — the per-epoch satellite candidate list, built once at rtk.cpp:892 and shared by the float filter and the ambiguity search. The mask is applied there, so both stay consistent.
  • RTKObs — column 15 of .pos, count of observations actually entering the RTK update. The real proof of effect: 70.60 → 21.00.
  • NumSat — column 10, the SPP-side satellite count. Does not react to NLOS exclusion, so don't judge the mask by it.
  • RTKLIB / demo5 / MRTKLIB / gici-open / claslib — other open GNSS packages this library benchmarks itself against in the README.
  • PPC — a Japanese positioning competition dataset (Tokyo, Nagoya runs) used as the main benchmark here. UrbanNav — Hong Kong/Tokyo urban-canyon dataset. smartLoc — German urban dataset.
  • Fisheye visibility — src/visibility/, an alternative NLOS source: point an upward fisheye camera at the sky, threshold the image against an open-sky reference, then project each satellite's az/el onto the image to decide LOS or NLOS. Same output as the mask, different sensor.
Back to top