Perception Robustness in Airside Adverse Conditions

How Airport-Specific Environmental Challenges Degrade Sensors and How to Compensate

1. Airside-Specific Conditions Not Found in Road Driving

Condition	Frequency	Severity for Sensors	Road Driving Equivalent
De-icing spray	Seasonal (daily in winter)	Critical — lens contamination	No equivalent
Jet engine exhaust	Every departure	High — thermal distortion	No equivalent
Jet blast debris	During engine run-up	Medium — physical impact risk	No equivalent
Standing water on apron	After rain	High — mirror reflections	Puddles (much smaller scale)
Tropical downpour (50mm/h)	Daily (Changi-type)	High — sensor blindness	Heavy rain (usually less intense)
Heat shimmer from tarmac	Summer afternoons	Medium — camera distortion	Highway heat shimmer (similar)
Apron lighting at night	Every night	Medium — harsh shadows, glare	Street lighting (more uniform)
Sun glare on wet tarmac	After rain + sunshine	High — specular reflection	Sun glare on wet road (similar)
De-icing fluid on surface	Winter operations	Medium — reduced LiDAR intensity	No equivalent
Snow/ice on apron	Winter	High — obscured markings	Snow on road (similar)

2. De-Icing Operations

2.1 Impact on Sensors

De-icing uses propylene glycol (Type I) and thickened glycol (Type IV) sprayed at 60-80°C:

LiDAR impact:
  - Glycol droplets on lens → scattered/blocked beams → blind spots
  - Glycol mist in air → backscatter → phantom returns (like fog)
  - Estimated degradation: 40-70% point cloud density loss in spray zone
  - Recovery: requires physical lens cleaning (not automatic wipers)

Camera impact:
  - Glycol streaks on lens → blurred/distorted images
  - Colored glycol (orange Type I, green Type IV) → color distortion
  - Warm glycol creates steam → temporary fog around aircraft

Radar impact:
  - Minimal — glycol droplets too small to scatter 77GHz waves
  - 4D radar provides reliable perception through de-icing operations

IMU/GPS impact:
  - None — unaffected by glycol

2.2 Mitigation Strategies

1. Operational avoidance:
   - Do not route autonomous GSE through active de-icing spray zones
   - Use ADS-B + de-icing schedule to predict and geofence zones
   - De-icing typically occurs at remote de-icing pads, not at stands

2. Sensor protection:
   - Hydrophobic lens coatings (repel glycol droplets)
   - Heated lens elements (prevent glycol from adhering)
   - Compressed air cleaning (periodic bursts to clear droplets)
   - Enclosed sensor pods with automatic cleaning

3. Graceful degradation:
   - Detect glycol contamination: sudden drop in LiDAR point count
   - Switch to 4D radar primary perception
   - Reduce speed to minimum safe level
   - Request remote operator intervention if degradation > 50%

4. Post-de-icing recovery:
   - Trigger sensor cleaning cycle after de-icing event ends
   - Verify sensor performance before resuming autonomous operation
   - Log incident for data collection (rare adverse condition)

3. Jet Engine Exhaust

3.1 Thermal Effects on LiDAR

Jet exhaust temperature: 300-600°C (idle), 1000-1500°C (takeoff)
Air density gradient: dn/dT ≈ -1×10⁻⁶ per °C

Effects on LiDAR:
  - Beam refraction through thermal gradient → point position errors (1-5cm)
  - Thermal turbulence → fluctuating returns → increased noise
  - At 905nm: minimal absorption by exhaust gases
  - At 1550nm: slightly more absorption, but still minor

Practical impact:
  - Objects behind exhaust plume: position error increases
  - Objects IN exhaust plume: phantom/noisy returns
  - Effects decrease rapidly with distance from exhaust
  - Most significant within 20-30m behind engine at idle thrust

3.2 Camera Effects

Heat shimmer:
  - Refractive index variations cause image wavering
  - AT-Net, MPRNet neural de-shimmer approaches exist (research stage)
  - Practical: use higher frame rate + temporal averaging

Exhaust plume visibility:
  - Visible as heat haze, especially against dark backgrounds
  - Can be detected as motion anomaly in optical flow
  - Thermal camera sees exhaust plume directly (useful for hazard detection)

3.3 Mitigation

1. Jet blast hazard zones (from POC 4):
   - Pre-computed hazard zones keep vehicle away from exhaust path
   - ADS-B + aircraft type → engine position + thrust estimate → hazard cone

2. Temporal filtering:
   - Exhaust turbulence is high-frequency noise
   - Low-pass filter on LiDAR returns reduces noise
   - Multi-frame accumulation averages out turbulence

3. 4D radar unaffected:
   - 77GHz radar passes through exhaust plumes without distortion
   - Provides reliable backup perception in exhaust zones

4. Standing Water on Tarmac

4.1 The Mirror Reflection Problem

LiDAR on standing water:
  - Water surface acts as specular reflector at low angles of incidence
  - Near-perpendicular beams: reflect off water → phantom ground returns BELOW actual surface
  - Oblique beams: reflect off water → miss (no return at all)
  - Result: inconsistent ground plane, false obstacles below ground level

Camera on standing water:
  - Mirror reflections of sky, aircraft, buildings
  - Confused semantic understanding (reflection of aircraft ≠ actual aircraft)
  - Specular highlights from sun/apron lights

Radar on standing water:
  - Minor effect at 77GHz (water is partially reflective)
  - Multi-path returns possible but low intensity

4.2 Detection and Compensation

python

def detect_standing_water(pointcloud, ground_model):
    """Detect standing water from LiDAR characteristics."""
    # Water indicators:
    # 1. Ground returns lower than expected (mirror bounce)
    # 2. Reduced return intensity (specular reflection away from sensor)
    # 3. Missing ground returns in patches (total specular reflection)

    ground_diff = pointcloud.z - ground_model.expected_z(pointcloud.x, pointcloud.y)
    intensity = pointcloud.intensity

    # Points below expected ground with low intensity → likely water
    water_mask = (ground_diff < -0.05) & (intensity < 0.3 * median_intensity)

    # Missing return patches (no points where ground expected) → possible water
    coverage_grid = compute_ground_coverage(pointcloud)
    missing_patches = coverage_grid < expected_coverage * 0.3

    return water_mask, missing_patches

5. Tropical Rainfall (Changi-Relevant)

5.1 Quantitative Degradation

Published data from K-Radar dataset and research:

Rain Intensity	LiDAR Point Loss	LiDAR Range Loss	Camera Impact	4D Radar Impact
Light (<2.5 mm/h)	~10%	~5%	Minor	None
Moderate (2.5-10 mm/h)	~25%	~15%	Moderate (blur)	None
Heavy (10-25 mm/h)	~40%	~25%	Severe (visibility)	None
Torrential (25-50 mm/h)	~56%	~35%	Near-blind	Negligible
Extreme (>50 mm/h)	~70%+	~50%+	Blind	Minor

Changi experiences 50mm/h regularly. At this intensity, LiDAR loses ~70% of points. UISEE developed custom rain filtering algorithms specifically for this.

5.2 UISEE's Rain Approach at Changi

What UISEE does (from deployment reports):
  - Custom LiDAR rain filtering algorithms (specific to Hesai XT)
  - Multi-frame temporal filtering (average across multiple scans)
  - Likely intensity-based rain point rejection
  - Maintained 20,000+ km accident-free including tropical conditions
  - Specific technical details not publicly disclosed

5.3 Mitigation Stack

Layer 1: Point cloud rain filtering
  - Statistical Outlier Removal (already in reference airside AV stack: airside_rain_detection)
  - Intensity-based rejection (rain drops have characteristic low-intensity returns)
  - Multi-return analysis (rain hits first, ground hits second)

Layer 2: Multi-frame accumulation
  - Accumulate 3-5 frames → rain points are random, real points are consistent
  - Moving average filter on occupancy grid

Layer 3: 4D radar as primary in heavy rain
  - Switch from LiDAR-primary to radar-primary when rain intensity > 25 mm/h
  - 4D radar provides reliable detections through all rain

Layer 4: Operational limits
  - If ALL sensors degraded > 60% → reduce speed to minimum
  - If degradation > 80% → controlled stop, wait for conditions to improve
  - UISEE operates through 50mm/h → achievable with proper filtering

6. Night Operations

6.1 Sensor Performance at Night

LiDAR: UNAFFECTED by lighting conditions
  - Active sensor (emits own light at 905/1550nm)
  - Same performance day and night
  - This is LiDAR's primary advantage for airside

Camera: SIGNIFICANTLY DEGRADED
  - Apron lighting is non-uniform (bright at stands, dark between)
  - High-pressure sodium lights (orange cast) → poor color perception
  - LED apron lights (new installations) → better color, harsh shadows
  - Aircraft navigation lights → potential confusion with markings
  - Exposure challenges: bright areas + dark areas in same frame

4D Radar: UNAFFECTED
  - Active sensor, works identically day and night

Thermal/LWIR: IMPROVED at night
  - Better contrast between warm objects (people, engines) and cool background
  - No sun glare interference
  - Personnel detection BETTER at night with thermal camera

6.2 Recommendation for Night

LiDAR-primary stack (the reference airside AV stack's current approach):
  → Night operations are inherently supported
  → No additional hardware needed for basic night capability

When cameras are added (Phase 2):
  → HDR mode essential for handling apron lighting variation
  → Consider global shutter cameras (less motion blur)
  → DINOv2 features are more robust to lighting changes than supervised backbones

For personnel safety at night:
  → Thermal/LWIR camera (FLIR Tura, ASIL-B rated)
  → 4x detection range vs headlights
  → Detects ground crew through hi-vis paradox (hi-vis causes 84-88% AEB failure at night)
  → UWB personal transponders as backup (10-30cm accuracy, lighting-independent)

7. Sensor Degradation Detection

7.1 Real-Time Health Monitoring

python

class SensorHealthMonitor:
    def check_lidar_health(self, pointcloud):
        """Detect LiDAR degradation in real-time."""
        metrics = {
            'point_count': len(pointcloud),
            'mean_intensity': pointcloud.intensity.mean(),
            'max_range': pointcloud.range.max(),
            'coverage_ratio': self.compute_coverage(pointcloud),
        }

        # Compare to baseline (learned from good conditions)
        degradation = {}
        for key, value in metrics.items():
            baseline = self.baselines[key]
            degradation[key] = 1.0 - (value / baseline)

        # Overall degradation score (0 = healthy, 1 = fully degraded)
        overall = max(degradation.values())

        # Determine probable cause
        if degradation['point_count'] > 0.3 and degradation['mean_intensity'] > 0.3:
            cause = 'rain_or_fog'
        elif degradation['point_count'] > 0.5 and degradation['mean_intensity'] < 0.1:
            cause = 'lens_contamination'  # blocked beams but remaining are normal intensity
        elif degradation['max_range'] > 0.3:
            cause = 'fog_or_visibility'
        else:
            cause = 'unknown'

        return SensorHealth(
            degradation=overall,
            cause=cause,
            action=self.get_action(overall),
        )

    def get_action(self, degradation):
        if degradation < 0.2:
            return 'NORMAL'
        elif degradation < 0.4:
            return 'REDUCE_SPEED'
        elif degradation < 0.6:
            return 'SWITCH_TO_RADAR_PRIMARY'
        elif degradation < 0.8:
            return 'MINIMUM_SPEED'
        else:
            return 'CONTROLLED_STOP'

8. Summary: Sensor Selection for Airside Robustness

Condition	LiDAR	Camera	4D Radar	Thermal	Recommended Primary
Clear day	Excellent	Excellent	Good	Fair	LiDAR
Clear night	Excellent	Poor	Good	Excellent	LiDAR + Thermal
Light rain	Good	Fair	Excellent	Good	LiDAR
Heavy rain	Poor	Very Poor	Excellent	Good	4D Radar
Fog	Poor	Very Poor	Excellent	Fair	4D Radar
De-icing spray	Very Poor	Very Poor	Excellent	Fair	4D Radar
Jet exhaust	Fair	Poor	Excellent	Excellent	4D Radar + Thermal
Standing water	Fair	Poor	Fair	Good	LiDAR + Radar
Snow	Fair	Poor	Excellent	Good	4D Radar
Sun glare	Excellent	Very Poor	Good	N/A	LiDAR

Key insight: 4D radar is the most critical sensor for airside robustness. It should be treated as a primary perception input, not a backup. The $50-200 cost per unit is negligible compared to the robustness it provides.

Sources

K-Radar dataset weather performance data
UISEE Changi deployment reports
Bijelic et al. "Seeing Through Fog Without Seeing Fog." CVPR, 2020
Heinzler et al. "Weather Influence and Classification with Automotive LiDAR Sensors." IV, 2019
RoboSense RSHELIOS IP67/IP6K9K specifications
Continental ARS548 specifications
FLIR Tura thermal camera specifications
IIHS AEB night testing results (hi-vis paradox)

SLAM Methods

Methods

Perception Robustness in Airside Adverse Conditions ​

How Airport-Specific Environmental Challenges Degrade Sensors and How to Compensate ​

1. Airside-Specific Conditions Not Found in Road Driving ​

2. De-Icing Operations ​

2.1 Impact on Sensors ​

2.2 Mitigation Strategies ​

3. Jet Engine Exhaust ​

3.1 Thermal Effects on LiDAR ​

3.2 Camera Effects ​

3.3 Mitigation ​

4. Standing Water on Tarmac ​

4.1 The Mirror Reflection Problem ​

4.2 Detection and Compensation ​

5. Tropical Rainfall (Changi-Relevant) ​

5.1 Quantitative Degradation ​

5.2 UISEE's Rain Approach at Changi ​

5.3 Mitigation Stack ​

6. Night Operations ​

6.1 Sensor Performance at Night ​

6.2 Recommendation for Night ​

7. Sensor Degradation Detection ​

7.1 Real-Time Health Monitoring ​

8. Summary: Sensor Selection for Airside Robustness ​

Sources ​