112. GPS Quality Filter

The GPS quality filter is a single-trajectory GPS-noise detector. It takes an already-decoded Ego trajectory — per-frame timestamps, WGS84 lon/lat, and a per-frame heading (yaw) — and returns a PASS/FAIL verdict plus supporting metrics and flags. It requires no map, running on input data only, making it inexpensive enough to run as a gate before any further more expensive steps.

112.1 Design goals

1. High precision — every rejection must be defensible to a human reviewer (target ≥ 90%).
2. Free recall — catch GPS-caused bad maps; accept that MapGen-internal bad maps (clean GPS → bad map) are out of scope.
3. Cheap — runs on the Ego-pose stream alone, so it can gate the pipeline early.

112.2 Detection method

The filter processes each clip as follows.

1. Local projection

   Project lon/lat to local ENU metres around the clip centroid (lat0 = mean(lat)):

   x = (lon − lon[0]) · 111320 · cos(lat0_rad)      # eastward, m
   y = (lat − lat[0]) · 111320                      # northward, m
   

   Flat-earth approximation; error ≪ 1 cm for clip extents under 2 km.

2. Directional residual

   For each interior frame i (not an endpoint, with finite yaw), take the residual of its position from the midpoint of its two neighbours, then project that residual onto the vehicle heading axes:

   mid     = (P[i−1] + P[i+1]) / 2
   r_vec   = P[i] − mid
   r_lon(i) = r_vec · (cos yaw, sin yaw)     # longitudinal — along heading
   r_lat(i) = r_vec · (−sin yaw, cos yaw)    # lateral — perpendicular, left-positive
   

   • Lateral residual — corrupts MapGen perception by shifting lane offsets and object positions. This is what the gate fails on.
   • Longitudinal residual is speed wobble along the path — cosmetic for MapGen and informational only.
   • The decomposition is invariant under rigid rotation/translation of the trajectory — which is exactly why the algorithm generalizes: any parser that can supply positions + heading can reuse it unchanged.

3. Usable-frame mask

   Only frames where the Ego is moving (speed > min_speed_m_s, default 5m/s) and on a straight section (|yaw_rate| < max_yaw_rate_deg_s, default 5°/s) are counted. On turns and while the Ego is parked, curvature and low speed inflate the GPS error estimate artificially.

4. Per-clip aggregation

   On the usable frames, the filter computes the median (p50) and 95th-percentile (p95) of |r_lat| and |r_lon| (reported in cm), plus the dropout percentage invalid_pct.

5. 1 Hz GPS-injection structure

   GPS receivers natively produce 1 Hz fixes; a 10 Hz stream is usually IMU-aided. When the raw 1Hz fix is injected once per second over IMU dead-reckoning, one frame-index-mod-10 phase carries a distinct error pattern. Detection runs on the total position error, not the directional one. The filter confirms this with two independent tests:

   • Phase test — phase_ratio = std(worst-phase residuals) / median (std of the other phases).
   • Spectral test — fft_ratio = |rFFT| at the 1 Hz bin / median background magnitude.

   Both must exceed their thresholds, and the amplitude must be non-trivial, for the test to fire.

   Note

   Clips with fewer than min_valid_frames (default 20) valid frames receive an immediate PASS with an info note — no statistics are computed.

112.3 Verdict logic and thresholds

A FAIL is produced by any severity flag; otherwise PASS. Info flags never fail a clip.

112.3.1 Severity flags

112.3.2 Info flags

• dropout (N%)
• 1Hz-structure (low amplitude)
• lateral-outliers (isolated spikes without sustained baseline)
• longitudinal (cosmetic)

112.4 Threshold calibration

Thresholds were calibrated on 56 manually tagged runs (24 good and 32 bad). Each gate sits at or beyond the p95 of the relevant Good-run metric.

112.4.1 Confusion matrix

                 DetectorRejects
TaggedBad     False   True
False            23      1       ← 24 Good
True             21     11       ← 32 Bad

Recall plateaus at approximately 35% because the 21 missed Bad runs all have clean GPS (lat_p95 < 30 cm) — they fail inside MapGen (lane fragmentation, connectivity holes, Ego off-map). A GPS-quality filter cannot flag those; catching them requires a complementary detector on MapGen-output topology, which is out of scope.

112.5 GPS quality filter configuration

All keys are nested under a gps_quality: sub-dict in the parser config: block. The full block is documented (commented, disabled) in nvidia_plugins/parsers/rclog_intervals_parser.yaml.

• Control

  Key	Default	Purpose
  enabled	false	Master on/off. The gate is a no-op (SKIPPED) when false.
  action	fail	fail — reject the clip and fail the pipeline stage; warn — log and write metrics, never reject.
  fail_on_flags	all severity flags	Optional allow-list of fail-worthy flags. Any severity flag not listed is downgraded to info.

  When a clip is rejected, the error message includes the path to a metrics file with the full quality report.

• Lateral residual gates (metres) — the primary tuning surface

  Key	Default	Raise to…
  lat_spike_p95_m	0.60	Tolerate larger single-frame lateral spikes
  lat_noisy_p50_m	0.05	Tolerate a higher sustained-noise median floor
  lat_noisy_p95_m	0.30	Tolerate a higher sustained-noise tail
  lat_1hz_amp_gate_m	0.20	Require larger 1 Hz amplitude before failing

• 1vHz injection confirmation

  Key	Default	Purpose
  phase_ratio_min	4.0	Worst-phase / median-of-others std ratio
  fft_ratio_min	3.0	FFT-peak / background ratio at the injection frequency
  injection_period_hz	1.0	Expected raw-fix injection frequency
  sample_rate_hz	10.0	Native GPS rate (drives phase period and FFT bins)

• Coverage, frame filtering and clip requirements

  Key	Default	Purpose
  invalid_pct_dropout	5.0	Dropout % that, combined with lateral noise, fails the clip
  min_speed_m_s	5.0	Exclude slow/parked frames (unreliable at low speed)
  max_yaw_rate_deg_s	5.0	Exclude turns (curvature inflates the GPS error estimate)
  min_valid_frames	20	Below this, short-circuit to PASS

112.6 Known limitations

112.7 Testing and verified behavior

112.8 GPS quality filter source files