Perception and Sensing for Autonomous Vehicles Under Adverse Weather Conditions: A Survey

Yuxiao Zhang; Alexander Carballo; Hanting Yang; Kazuya Takeda

Source-linked AI summary

Perception and Sensing for Autonomous Vehicles Under Adverse Weather Conditions: A Survey

Yuxiao Zhang, Alexander Carballo, Hanting Yang, Kazuya Takeda

arXiv:2112.08936v2Dec 16, 2021cs.RO

TL;DR

Adverse weather degrades ADS sensor perception, a central limitation for autonomous driving. This survey reviews sensor effects and mitigation approaches, finding progress in common wet weather while snow, strong light, and contamination remain less developed.

Problem
Adverse weather can obscure lane marks and affect sensors, challenging lateral localization needed to keep vehicles within lanes.
Source: §7.4 · p. 27
Method
The survey reviews adverse-weather effects on major ADS sensors and AV elements, perception-enhancement methods, and hardware, auxiliary, and future-sensor solutions.
Source: §12 · p. 39
Results
Performance has progressed in common wet weather, while snow remains at the dataset-expansion stage and strong light and contamination have few research solutions.
Source: §12 · p. 39
Takeaways & Limitations
The survey concludes that research is tending toward robust sensor fusion, sophisticated networks and computer-vision models, and greater legislative and public-acceptance support.
Source: §12 · p. 39
Perception degradation is the main limitation of ADS sensors in adverse weather, while sensor temperature and humidity endurance are barely analyzed systematically.
Source: §11.3 · p. 36

Abstract

from arXiv · show

Automated Driving Systems (ADS) open up a new domain for the automotive industry and offer new possibilities for future transportation with higher efficiency and comfortable experiences. However, autonomous driving under adverse weather conditions has been the problem that keeps autonomous vehicles (AVs) from going to level 4 or higher autonomy for a long time. This paper assesses the influences and challenges that weather brings to ADS sensors in an analytic and statistical way, and surveys the solutions against inclement weather conditions. State-of-the-art techniques on perception enhancement with regard to each kind of weather are thoroughly reported. External auxiliary solutions, weather conditions coverage in currently available datasets, simulators, and experimental facilities with weather chambers are distinctly sorted out. Additionally, potential future ADS sensors candidates and approaches beyond common senses are provided. By looking into all kinds of major weather problems the autonomous driving field is currently facing, and reviewing both hardware and computer science solutions in recent years, this survey points out the main moving trends of adverse weather problems in autonomous driving, i.e., advanced sensor fusions, more sophisticated networks, and V2X & IoT technologies; and also the limitations brought by emerging 1550 nm LiDARs. In general, this work contributes a holistic overview of the obstacles and directions of ADS development in terms of adverse weather driving conditions.

1. Introduction

Adverse weather remains a major barrier to reliable autonomous driving, despite the promise of AVs to improve safety, efficiency, and mobility. This survey examines weather’s effects on ADS and reviews existing solutions, validations, and future research directions.

Introduction

AVs promise fewer crashes, lower energy use and emissions, and greater mobility access, but adverse-weather performance remains a pressing ADS problem.The introduction also notes that casualties involving ADS-equipped vehicles are rising.Source: §1 · p. 1
11.0% of global time has precipitation, rain raises accident risk 70%, 77% of countries receive snow, and U.S. winter roads see over 30,000 annual crashes.Fog, haze, sandstorms, and strong light reduce visibility, while heat, cold, contamination, and hardware damage can also affect vehicles.Source: §1 · p. 1
Current AVs still face weather-related operational limits: some stop in heavy rain or snow, while low temperatures can impair charging and reduce operating hours.The Mcity shuttle is shut down when its wipers must run continuously; Sensible4 has begun snowy-weather test drives. Level 2 systems also struggle with slippery or low-visibility conditions, and Super Cruise prohibits use in rain, sleet, fog, ice, or snow.Source: §1 · p. 2
Existing reviews do not comprehensively cover all weather phenomena, direct and secondary AV impacts, component performance, and hardware and software solutions.The authors identify this as a gap their survey aims to address.Source: §1 · p. 2
The survey presents statistical analyses of weather’s effects on ADS sensors, reports sensor-fusion and mechanical solutions and perception-enhancement algorithms, and conducts experimental validations.It also proposes future research directions and discusses trends and limitations in adverse-weather autonomous driving.Source: §1 · p. 2

2. Overview of Autonomous Vehicles

AVs use sensors to perceive their surroundings and navigate with little or no human input, but weather can disrupt sensing and downstream driving functions. Direct effects on the vehicle and road can feed back through vehicle behavior and surrounding traffic, forming a cycle.

2. Overview of Autonomous Vehicles

Weather changes environmental states and sensor perception, complicating detection, planning, and control while wind and road conditions can directly alter vehicle behavior.Changes in ego-vehicle and surrounding-vehicle states can in turn affect sensor operation and the environment, creating a feedback cycle.Source: §2 · p. 3, §2.3 · p. 5

2.1.1. LiDAR

LiDAR is a core, high-accuracy, illumination-independent AV sensor, with return selection and power/noise adjustments that can help in denser fog.LiDAR characteristics include range, accuracy, point density, scan speed, configuration, wavelength, environmental robustness, form factor, and cost; in denser fog, the last return can overlap the reference point cloud more than the strongest return, while higher laser power extends range.Source: §2.1.1 · pp. 3–4

2.1.2. Camera

Cameras are indispensable but highly vulnerable in adverse weather, ranging from windshield-mounted dashcams to professional and fisheye camera sets used for broader data collection.Dashcams can record with a field of view as wide as 170°; night-vision cameras are discussed as specialized options.Source: §2.1.2 · p. 4

2.1.3. Radar

Automotive radar uses reflected radio waves, typically at 24–77 GHz and sometimes 122 GHz, to measure object distance, speed, and direction.FMCW radar derives range from the frequency difference between transmitted and reflected signals and offers superior range resolution and accuracy in autonomous driving.Source: §2.1.3 · p. 4

2.1.4. Ultrasonic

Ultrasonic sensors are inexpensive, commonly support parking and blind-spot assistance, and normally detect no farther than 11 m, limiting them chiefly to close-range uses [29, 30].They operate around 40–70 kHz; Tesla’s summon feature uses them to navigate parking spaces and garage doors.Source: §2.1.4 · p. 4

2.1.5. GNSS/INS

GNSS uses L-band signals that pass through clouds and rain with minimal path attenuation, while GNSS+INS and other sensor combinations can improve positioning accuracy and reduce drift.GNSS may be combined with RTK corrections; INS uses dead reckoning from motion, rotation, and magnetic-field sensors. Non-GNSS options such as Wi-Fi, Bluetooth, UWB, landmarks, V2I, and RFID support indoor, tunnel, or urban positioning. A cited Spatial FOG Dual GNSS/INS system reports 8 mm horizontal accuracy and about 0.005° roll/pitch accuracy.Source: §2.1.5 · pp. 4–5

2.2. Object detection & tracking and localization

Perception supports safe navigation by detecting objects and their locations; tracking predicts other parties’ trajectories, while localization estimates the AV’s position relative to a reference frame.Vision systems commonly label objects with bounding boxes, and YOLO formulates detection as regression; LiDAR point-cloud methods include PointPillars, SECOND, and Voxel-FPN. Tracking uses heading and velocity to predict trajectories, but the paper leaves detailed tracking treatment outside its weather-focused scope.Source: §2.2 · p. 5

2.3. Planning and Control

Planning selects a global route and a local path that avoids obstacles and rule violations, while steering, throttle, and brake inputs execute the AV’s decisions.Weather-related perception disruption complicates downstream planning and control; ECU or VCU safety measures can also use simple sensors such as sonar.Source: §2.3 · p. 5

3. Adverse Weather Influences

Weather degrades ADS sensing through direct effects on sensor signals and images, as well as indirect effects from snow, spray, wind, and road conditions. LiDAR, radar, cameras, ultrasonic sensors, and GNSS respond differently, with important limitations and potential sources of redundancy.

3.1.1. Rain and fog

Ordinary rain has limited direct LiDAR signal attenuation, but emitter-window wetting, heavy rain, and fog can reduce reliability and produce noisy or misleading point clouds.Small-aperture LiDAR tests found losses above 10% very unlikely under a 99.5% signal criterion, while emitter-window wetting caused attenuation from 25% to about 50%. LIBRE tests detected highlighted objects 13 m ahead in dense fog, 47 m in lighter fog, and 24 m in rain; sprinkler-generated water was also detected as vertical cylinders.Source: §3.1.1 · pp. 6–7, Figure 6 · p. 7

3.1.2. Snow

Snow swirl can create voids and shorten LiDAR view distance; low temperatures can exceed operating limits and reduce ranging precision, while aerial LiDAR under turbulence remains understudied.A temperature shift from −20°C to +60°C increased measurement delay by about 6.8 ns, widening range by over 1 meter. Satellite and single-photon LiDAR can look through haze or sandstorms, but aerial systems face turbulence-related scintillation, beam spreading, and wander.Source: §3.1.2 · pp. 7–8, §3.1.3 · p. 8

3.2. Influence on radar

At 77 GHz, radar attenuation is lower than 905 nm LiDAR’s under the stated heavy-rain conditions, and radar retains useful detection in fog, rain, and snow despite limited spatial resolution.The cited attenuation levels are 10 dB/km for radar and about 35 dB/km for LiDAR. Simulated 77 GHz radar attenuation in 10 mm/hr snowfall was about 6 dB/km; radar can still struggle with pedestrian detection and object shape and size classification.Source: §3.2 · pp. 8–9

3.3.1. Rain and fog

Rain drops on a camera lens can block or distort images, fog deprives cameras of visual information, and a camera-in-the-loop study found up to a 40% rise in miss rate at night or in fog.The study’s closed-loop setup evaluated object detection under varying weather conditions.Source: §3.3.1 · p. 9

3.3.2. Snow

Snow or hail can obscure image edges and cause detection failures, while accumulated snow and ice can hide lane markings and compromise camera data acquisition.Snowflakes that melt and refreeze on the camera window can form an opaque blockage.Source: §3.3.2 · pp. 9–10

3.3.3. Light conditions

Strong illumination can nearly eliminate camera visibility and create glare, while reflections from glossy surfaces or windshields can supply false information that a 2D camera cannot readily identify.HDR cameras can mitigate visibility drops from abrupt light changes, including tunnel entry and direct sunlight.Source: §3.3.3 · p. 10

3.4. Miscellaneous issues

Flying debris can damage sensor windows, surface contamination can cut LiDAR range, and spray from nearby vehicles can cause ghost targets, blockage, false positives, and missed detections.Tests with near-homogeneous dust on a scanner surface showed a 75% reduction in LiDAR maximum range; water-spray data have also been augmented for classifier training.Source: §3.4 · pp. 10–11
Ultrasonic sensing can add close-range redundancy in low visibility and help classify road surfaces, while GNSS signals are attenuated by water, with wind increasing rain-related attenuation.Rain below 30 mm/hr yielded about 96% L-band transmissivity; a 1 mm liquid-water layer caused GPS signal loss over 9.4 dB, compared with 0.9 dB through 4 cm of snow and 1.7 dB through 14 cm.Source: §3.4 · pp. 10–11

4. Sensor Fusion and Mechanical Solutions

The section surveys weather-agnostic sensor fusion and mechanical mitigations: multimodal sensing can outperform individual sensors, while protection, cleaning, passive accessories, and recalibration address weather-related sensor challenges.

4.1. Sensor fusion modalities

Fusion of LiDAR, camera, radar, and weather-related sensors can outperform individual sensors in adverse weather and overall perception; synchronization and calibration remain important.Monrroy Cano et al. developed a toolbox with centimeter-level extrinsic calibration for LiDAR-to-LiDAR and camera-LiDAR fusion, integrated into Autoware.Source: §4.1 · p. 11, Figure 10 · p. 11

4.1.1. Radar dominant

Radar-LiDAR RadarNet showed rain robustness for long-distance detection and velocity estimation; radar-camera fusion beat LiDAR-only detection by over a third but struggled with horizontal-target tracking.Fritsche et al. fused a fog-penetrating mechanical pivoting radar with LiDAR for landmark search and SLAM, but tested it on a robot rather than an AV.Source: §4.1.1 · pp. 11–12

4.1.2. Camera dominant

The LWIR-based AEB sensor suite avoided hitting any mannequin in adverse tests, while normal AEB almost hit every one; LWIR also performed better in thick fog than MWIR and SWIR.Other camera-dominant approaches combine RGB and NIR for low-light imaging or fuse visible and thermal imagery for segmentation [95, 96]; thermal cameras’ lower operating-temperature bounds may constrain winter use.Source: §4.1.2 · p. 12

4.1.3. Comprehensive fusions

Weather-adaptive sensor weighting can improve perception robustness, and Bijelic et al. [98] reported better fog detection than LiDAR- or image-only sensing, though their many-sensor system burdens vehicle power and response time.Robust-SENSE [97] uses supplementary sensors when LiDAR performance degrades; the section emphasizes real-time sensor-selection strategies and weather-specific weighting as determinants of reliability.Source: §4.1.3 · pp. 12–13

4.2.1. Protection and cleaning

Sensor housings, windshield placement, and wipers protect against snow, rain, frost, contamination, and impacts; heaters and ultrasonic cleaning can remove ice, mist, dirt, and debris.Housing sensors behind glass is not suitable for LWIR cameras, and automatic contamination detection is still developing; ultrasonic cleaning uses 35 kHz vibration to transfer dirt into a fluid film.Source: §4.2.1 · p. 13

4.2.2. Passive accessories

Hydrophobic membranes improve visual distance and decrease minimum visual angle by almost 34% [120]; KoPro’s Musou Black flock sheet reflects less than 0.5% in visible light and NIR for about US$30 [107].The hydrophobic treatment also benefits object detection and road-sign reading [105]; the flock sheet reduces cabin reflections and improves perception in sunlight, as shown in Figure 12.Source: §4.2.2 · pp. 13–14, Figure 12 · p. 14

4.2.3. Integrity and re-calibration

Mechanical damage, aging, and changes in mounting or vehicle height can misalign ADS sensors, motivating durable components, structural-integrity alerts, and sensor recalibration.Recalibration may be static or dynamic and can be costly; Apple’s infrared windshield-chip detection patent illustrates an approach to identifying shield damage early.Source: §4.2.3 · p. 14

5. Perception Enhancement Methods and Experimental Validations in Different Weather Conditions · 5.1. Rain

This section surveys perception enhancement for adverse weather, with rain research spanning vehicle-based precipitation sensing, LiDAR performance quantification, and camera de-raining. It also identifies limitations in current de-raining methods, particularly for adherent drops, heavy rain, and dynamic scenes.

5. Perception Enhancement Methods and Experimental Validations in Different Weather Conditions

The section focuses on object detection as a key ADS capability to enhance against adverse weather and organizes perception-enhancement research by weather type.Table 3 indexes the surveyed solutions by weather type.Source: §5 · p. 14

5.1. Rain

Vehicle-based rain sensing includes simulated wet/dry observations, LiDAR rainfall-rate estimation in a weather chamber [127], and connected-car wiper-based rainfall mapping [125].The earlier wet/dry methods were validated against weather radar but not real-world data; a commercial in-vehicle weather station was considered impractical [98].Source: §5.1 · p. 15
Filgueira et al. [126] quantified rain-related LiDAR range, intensity, and point-count changes, while Goodin et al. [124] modeled performance from rain rate and clear-condition maximum range.Filgueira et al.’s roadside test had orientation-related impartiality concerns; Goodin et al.’s simulation-derived equation had no field validation, though the paper describes quantifiable thresholds as useful for decisions.Source: §5.1 · p. 15
Camera de-raining addresses falling and adherent raindrops; Quan et al. [133] fused streak and raindrop removal with attention, while using a real-world rain dataset.Quasi-Sparsity-based training [130] and continual learning [131] were also applied to rain-streak removal.Source: §5.1 · p. 15
RIC-Net [132] combines background extraction, rain-streak elimination, and intensity control; trained from drizzle to downpour, it showed superiority in real-data validation.Histogram of oriented gradient and auto-correlation losses support orientation consistency and suppression of repetitive streaks.Source: §5.1 · pp. 15–16
Rain-generation methods support removal training: Wang et al. [134] built a Bayesian image generator, Ye et al. [129] used disentangled image translation, and Yue et al. [128] developed semi-supervised video de-raining.Yue et al.’s dynamic generator models rain-streak spatial and temporal parameters with emission and transition components.Source: §5.1 · p. 16
Many de-raining algorithms perform poorly with adherent drops, higher rain rates, or dynamic scenes, and real-time adherent-drop removal incurs processing latency.The paper also notes that PSNR and SSIM improvements are not promising for raindrop detection and removal compared with deep-learning and CNN methods.Source: §5.1 · p. 16

5.2. Fog

Fog degrades LiDAR point clouds and camera images, prompting waveform-based sensing, sensor fusion, and image de-hazing approaches. The section also reports increased pedestrian detections after GAN-based de-hazing and notes persistent challenges in real-world training data and evaluation.

5.2.1. Fog in point clouds

Waveform identification and multi-echo processing enable LiDAR to distinguish fog returns and filter false targets, but detection can fail when visibility falls below 10 m.Pfennigbauer demonstrated classification within 30 m; SICK AG’s HDDM+ technology recognizes multiple environmental waveforms. Full-waveform and single-photon approaches were also explored, with 1550 nm manufacturing difficulties noted.Source: §5.2.1 · p. 16
LiDAR–radar fusion outperformed LiDAR alone in fog, while LiDAR–camera SLS-Fusion improved performance by over a quarter after fog reduced its original performance by almost half.MVDNet [139] fuses region-wise features from LiDAR and radar; SLS-Fusion was trained on clear and foggy data.Source: §5.2.1 · p. 17

5.2.2. Fog in images

Image de-hazing uses either image-processing enhancement or atmospheric-scattering restoration, whose estimation is ill-posed because scene radiance, transmission, and atmospheric light are unknown.Enhancement can improve contrast and texture but may distort color and fails to exploit scene depth effectively.Source: §5.2.2 · p. 17
Deep de-hazing methods address domain shift, negative information, temporal redundancy, and UHD-image complexity through unsupervised fine-tuning, contrastive regularization, video alignment, and bilateral learning.These approaches include PSD, ACER-Net, CG-IDN, and a framework for 4K-resolution images.Source: §5.2.2 · pp. 17–18
Unpaired image translation can avoid the need for pixel-matched hazy and clear images, but de-hazing results often contain artifacts.Cycle-Dehaze combines cycle-consistency and perceptual losses; Shao et al. translate between real and synthetic haze domains.Source: §5.2.2 · p. 18
Real paired haze data are difficult to collect, so de-hazing evaluation relies on non-reference metrics; generated haze datasets are a potential route to better metrics and deployment.The section describes datasets made with a professional haze/fog generator and a multiple-weather stacking architecture.Source: §5.2.2 · p. 18

5.2.3. Gan-based de-hazing model experimental evaluation

Haze removal increased valid pedestrian detections, especially for people partially obscured in the background.The authors tested a CycleGAN-based model with a weather-layer loss and spatial feature transform on Cityscapes and Foggy Cityscapes data.Source: §5.2.3 · p. 18

5.3. Snow

The survey describes snow perception as both pathfinding over snow-covered roads and recovering degraded camera or LiDAR inputs during snowfall. It reviews filtering and sensor-fusion methods, alongside validation results and limitations.

5.3.1. Snow covering

Snow-cover pathfinding and segmentation use image filtering and sensor fusion, with thermal input complementing RGB and camera–LiDAR–radar streams delineating drivable areas.Yinka et al. report close to 100% pathfinding across 100 road-image frames in a simple roadside-snow scenario. Vachmanus et al. report improved snow segmentation over RGB-only and several state-of-the-art networks; Rawashdeh et al. describe filtering misleading road lines and edges to identify the general drivable area.Source: §5.3.1 · p. 18

5.3.2. Snowfall

For LiDAR snowfall, DROR adapts its search radius to preserve distant environmental points while removing nearby snow noise; LIOR uses intensity differences to retain more points than DROR.Charron et al. report nearly four times the precision of normal ROR filters, preserving points 6–18 m away while removing salt-and-pepper noise within 6 m. Park et al. note LIOR’s advantage depends on the circumstances.Source: §5.3.2 · p. 19, Figure 14 · p. 20
Camera de-snowing uses DDMSNet’s coarse removal followed by semantic and geometric priors, while synthetic snow generation aims to make training images more realistic.Zhang et al. evaluate DDMSNet on snowy images created from Cityscapes and KITTI, while noting the need for photo-realistic snow simulation. Von Bernuth et al. render flakes into reconstructed 3D scenes using depth, physical and meteorological principles, and motion blur.Source: §5.3.2 · pp. 19–20

5.3.3. Point cloud de-snowing validations

DSOR combines DROR’s range adaptation with SOR’s thresholding, matching DROR in de-noising and feature preservation while computing faster.The intensity filter gave an acceptable result but depends on a scene-specific interval, limiting practical use; the entropy filter still struggles in very dense snowfall.Source: Figure 14 · p. 20, §5.3.3 · p. 21, §5.4.2 · p. 21

5.4.2. Reflections and shadows

Reflection-removal techniques can relieve image ambiguity, but fail in extremely strong light; mirage detection has limited real-world validation, while shadows retain image elements at lower brightness.Kumar et al. detected a road mirage and estimated depth in video, but the lack of mirage effects in datasets makes accuracy difficult to validate. The supplied shadow discussion describes generating and removing shadows.Source: §5.4.2 · p. 21

5.4. Light related

Strong light and glare impair AV perception, while thermal imaging and image-enhancement methods can preserve or improve object detection and classification. The section also notes that LiDAR and radar mitigate threats from reflection and shadow in mature ADS.

5.4.1. Strong light and glare

Literature specifically addressing strong light and glare is limited; Maddern et al. (2014) used an illumination-invariant transform, and LiDAR and radar mitigate reflection and shadow in mature ADS.The transform removed almost all variation attributed to sunlight intensity, direction, spectrum, and shadow; the algorithm can also enhance low-light images.Source: §5.4.1 · p. 21

5.4.3. Thermal imaging validations

At 40 m from a 200 klx Xenon source, LiDAR and RGB yielded too few measurement points for recognition, while thermal imaging discerned the board beneath the light.The figure reports four mannequin points, three reflective-target points, and ten black-vehicle points; classification was not possible, while thermal-camera objects were clearly discernible.Source: §5.4.3 · p. 21, Figure 15 · p. 22
Multiscale Retinex restores image color and contrast, while the bio-inspired retina removes noise and illumination variation and enhances static and dynamic contours.Both methods were applied to thermal images to enhance perception.Source: §5.4.3 · p. 22
For person/car 1/car 2, YOLOv3 recall was 0/0/0% with RGB, 44.3/0/60% with thermal, 58.6/30/62.9% with multiscale Retinex, and 67.1/45/78.6% with bio-inspired retina.The comparison analyzed 10 seconds of camera frames while the car approached a strong light source.Source: §5.4.3 · p. 22

5.5. Contamination

Contamination can obstruct ADS sensor line of sight; the survey describes synthetic soiling-data methods, automatic cleaning controls, and a safety scanner designed to keep detecting nearby objects despite debris.

5.5. Contamination

Contamination can invade a sensor’s line of sight, as naturally formed mud on a Cadillac XT5 backup camera illustrates.The resulting robustness and adaptability demands are described as a rigorous test for ADS.Source: §5.5 · p. 22, Figure 17 · p. 23
Uřičář et al. developed DirtyGAN for region-specific synthetic lens soiling, extending SoilingNet’s opaque and transparent examples into a Woodscape-based Dirty dataset.CycleGAN generated random soiling patterns and masks; DirtyGAN restricted the mask area, and degradation evaluation on Cityscapes was reported as successful.Source: §5.5 · p. 22
Trierweiler et al. used total internal reflection to distinguish liquid from solid windshield contamination and determine whether to trigger a nozzle as well as a wiper.The device splits red-light diffuser intensity into two sections; nonuniform windshield curvature raises concerns about installation and operational consistency.Source: §5.5 · p. 23
SICK’s microScan 3 detects nearby objects without false alarms despite sawdust on its emitter cover, using a 275° scan, 845 nm lasers, and 0.1° angular resolution.The scanner uses HDDM+ and is presented as a potential final barrier when contamination compromises other sensors.Source: §5.5 · p. 23

6. Classification and Assessment

The section surveys local weather and visibility classification for autonomous vehicles and argues that adverse-weather decisions also need risk assessment that accounts for uncertain sensor inputs and reduced stopping sight distance.

6.1. Weather classification

Local weather classification complements perception enhancement, but camera-based methods remain unsaturated and ADS need finer, more precise classification that may extend beyond images.Methods progressed from few-class recognition to feature-based multiclass classifiers and deep CNNs; one tested ResNet-18 achieved rather low accuracy.Source: §6.1 · pp. 23–24
Multi-echo LiDAR classifies rain and fog using echo statistics, but limited field of view and dynamic scenes may weaken robustness; camera–LiDAR fusion is proposed as a direction.Heinzler et al. used kNN and SVM classifiers, while Dannheim et al. proposed fusion based on rain- and fog-induced backscattering intensity differences.Source: §6.1 · p. 24

6.2. Visibility classification

Visibility estimation for ADS mainly uses camera neural networks and range classes, while LiDAR backscatter estimation is practically limited and precipitation rates alone often do not reflect road visibility.Visibility crises can arise from water screens and mist during rain; direct precipitation-to-visibility correspondence is most relevant to thick fog, pollution smog, or sandstorms, and future IoT or V2X may support real-time classification.Source: §6.2 · pp. 24–25
The hybrid fog-visibility model achieved 85% or higher overall accuracy, enabling ADS to set thresholds for dangerous low visibility.Duddu et al. [215] combined block-level Shannon entropy with image features to classify visibility as 0–50 m, 50–150 m, or above 150 m.Source: §6.2 · p. 24

6.3. Risk assessment

Adverse weather can impair sensors or reduce visibility below designed stopping sight distance, increasing uncertainty; the paper calls for reliability analysis in visibility-related decisions and control.The section notes a 0.5 s assumed AV reaction time and argues that dense information at uncertain locations can help assess risk in advance.Source: §6.3 · pp. 25–26

7. Localization and Mapping

Localization in adverse weather relies on GPS-IMU, SLAM, a-priori maps, and sensor-based methods, each facing distinct robustness challenges. The survey describes approaches to maintain position estimates when weather changes or obscures road features.

7. Localization and Mapping

WhONet reduced positioning error by over 90% in wet and muddy, GNSS-deprived tests, but its performance under weather-induced road conditions remains unverified.GPS or INS alone generally has double-standard deviation above 1 meter on most roads, and the WhONet tests used the CUPAC dataset; the network targets tire changes rather than weather.Source: §7 · p. 26

7.1. Simultaneous localization and mapping

Sequence-based visual SLAM addresses seasonal and weather-driven appearance changes, but alignment assumptions, localization drift, and map insufficiency limit robustness and long-distance use.SeqSLAM matches coherent navigation sequences, while a later method uses DCNN descriptors and sequence information for loop closure; 4Seasons offers cross-weather data and accurate reference poses when GNSS reception is good.Source: §7.1 · p. 26

7.2. A-priori map

A-priori map localization searches for mapped landmarks, from snow poles that mark road edges to roadside signs, but maps require ongoing maintenance.Accumulated snow occludes features and poses a greater localization challenge than snowfall; RFID checkpoints and V2I also support localization, while Ford’s marker matching can locate AVs to centimeter accuracy without extra hardware cost.Source: §7.2 · pp. 26–27
LGPR localizes through subsurface radar reflections and can navigate puddles and snowdrifts, while surfel maps built from year-long LiDAR data model road features across seasons.LGPR achieves at least equal or better accuracy than traditional localization methods, but saturated soil can attenuate its signal by 10 dB/m and snow-melting salt also degrades it.Source: §7.2 · p. 27

7.3. Sensor solutions

Multisensory place recognition improves robustness; multiresolution LiDAR maps reduced lateral and longitudinal RSM error by around 80% in snowy conditions.Navtech radar supported map building and localization in rain and snow; the LiDAR approach had a small RSM error increase in normal downtown conditions compared with reflectivity-based localization.Source: §7.3 · p. 27
Combining visual and infrared SLAM reduces error versus either modality alone by assessing sensor data quality and suppressing affected inputs, including in smoke, fire, and extreme heat.The method was also tested in low visibility, dusk, and sudden strong artificial light; flawed data may still influence the algorithm.Source: §7.3 · p. 27

7.4. Lateral localization

Snow, rain, and obscured or noisy lane features can disrupt lateral localization; PCA edge-profile matching reduced lateral movement error to 15 cm with 96.4% accuracy in critical environments.The method removes false lane lines and reconstructs missing LiDAR profile elements using leading eigenvectors.Source: §7.4 · pp. 27–28

8. Planning and Control

Weather-aware planning and control must adapt routes, trajectories, speed, following distance, and steering to hazards, while integrating sensing, monitoring, and passenger communication.

8.1. Planning

Extreme-weather road damage can require global rerouting; Japan’s CERI Snow and Ice Research Team is developing weather alerts and snowstorm detours, requiring IoT or at least V2V.The service provides snowfall, visibility, and forecast information to drivers’ phones.Source: §8.1 · p. 28
Repeatedly driving identical shuttle lines can refine trajectories to centimeter-level precision yet create worn grooves or perilously slippery winter tracks.The section identifies winter driving patterns and road-wear prevention as local-planning concerns.Source: §8.1 · p. 28
Under limited visibility, Taş et al. plan robustly for worst cases; Jalalmaab et al. use time-varying MPC constraints, while Peng et al. adapt trajectories to rain.Taş’s planner slows and prepares to yield for unseen vehicles; Peng et al. use a Gaussian model of rain’s pavement impact and score-based decisions to maintain safety, comfort, and efficiency. The section notes these methods depend on weather classification or uncertainty assessment and integrated sensors and modules.Source: §8.1 · p. 28

8.2. Control

At 80 km/h, braking distance rises 28% in snow and 71% on ice, reaching 180 m for sedans and 300 m for trucks and buses.The section says longer following distances and, if needed, lower cruising speeds are required on wet or frozen roads.Source: §8.2 · p. 28
For crosswinds, ESC corrects lateral deviation after skidding begins, so ADS would benefit from proactive steering control with better judgment than a human driver.Crosswinds can exert a large force perpendicular to the vehicle’s heading, causing side-slip.Source: §8.2 · p. 29
Weather-related slipping and traction loss can undermine localization and control execution; Sensible4 monitors wheel and motion data, while remote operators can take over when necessary.The monitored data include actual acceleration, wheel rotation speed, and wheel angle; remote control is described as a final barrier for Level 4 or higher vehicles without safety pilots.Source: §8.2 · p. 29
Passenger comfort requires avoiding unnecessary acceleration and abrupt braking, while HUDs can build confidence by showing upcoming AV movements, especially in low visibility [257, 244].The section notes that distrust of AV behavior in adverse conditions can unsettle passengers.Source: §8.2 · p. 29

9. Auxiliary Approaches in Adverse Conditions

The section surveys auxiliary approaches that support adverse-weather driving through road-surface sensing, roadside infrastructure, smart headlights, and other tools for maintaining vehicle functions. It describes their capabilities alongside implementation limits and conditions.

9.1. Road surface detection

Road-surface sensing can aid weather classification: vision methods classify wet or dry conditions at 86% accuracy in autonomous driving, and DNNs estimate friction and pavement-type combinations.Road-surface changes, especially wetness, reflect weather; friction levels can correspond to clear, rain, snow, and freezing conditions, supporting vehicle dynamics control and ABS.Source: §9.1 · p. 29
Audio networks can detect road conditions from more than 780,000 audio bins, but vehicle or tire variation and sound-device installation may limit universality.The method can use sounds from passing vehicles even when the ego vehicle is stopped.Source: §9.1 · p. 29

9.2. Roadside Units

Roadside LiDAR can support vehicle perception and planning in wind and snow, while connected RSUs can share local weather classifications and perception data with nearby vehicles.Its elevated, wide-angle view avoids some rain or snow interference caused by surrounding vehicles and can provide redundancy when onboard sensors are unreliable.Source: §9.2 · p. 30
F-scores for roadside LiDAR weather classification normally exceed 90% for clear, fog, rain, and snow, while another study’s weather records excluded fog.The studies used weighted k-nearest neighbors on LiDAR parameters and point-cloud detection-distance offsets, respectively; the missing fog data reflected geographic limitations.Source: §9.2 · p. 30

9.3. Others

Smart headlights that avoid illuminating rain or snow particles improve visibility, and a camera–LiDAR headlight detected human-sized objects at 20 m with an 85% recognition rate.The visibility-improving headlights tracked particles and switched off intersecting light rays with few losses in light throughput; the authors suggest they could also help ADS cameras.Source: §9.3 · p. 30
OpenXC Android applications support adverse-condition assistance, including webcam-based nighttime collision warnings and LiDAR-based alerts when vehicles approach with excessive momentum.The platform reads vehicle metrics and supports integration with connected services.Source: §9.3 · p. 30
Other work targets weather-robust pedestrian detection, vehicle tracking, radar tracking, and lane tracking on unmarked roads, alongside pedestrian datasets and simulation data.Approaches include micro-Doppler spectra, multilayer deep learning, a Gaussian-mixture PHD-filter tracker, and color-based vision lane tracking.Source: §9.3 · p. 30

10. Tools

This section surveys adverse-weather datasets and the simulation and experimental facilities used to develop and test autonomous driving systems. It highlights limited weather diversity in common datasets and the safety and controllability benefits of simulated and laboratory testing.

10.1. Data sets

Most commonly used training datasets contain few conditions beyond clear weather; nuScenes includes some rain, A*3D rain at night, and ApolloScape strong light and shadow [297].The survey summarizes weather coverage and collection sensors for datasets in Table 6.Source: §10.1 · p. 31
CADC targets snow with eight cameras, LiDAR, and GNSS+INS, while Oxford RobotCar includes heavy rain, snow, direct sunlight, night, and road or building works [290].CADC’s large amount of snow supports testing detection, localization, and mapping across snow conditions that are difficult to reproduce artificially.Source: §10.1 · p. 32
Foggy Cityscapes adds fog synthesis to more than 20,000 clear-weather images, while ACDC provides pixel-level annotations for fog, nighttime, rain, and snow [300].ACDC supports training and testing semantic-segmentation methods on adverse visual conditions.Source: §10.1 · p. 32

10.2. Simulators and Experimental facilities

Simulators offer customizable scenarios and weather testing anytime and with zero risk, including adjustable rain, snow, fog precipitation rates, and strong light.They address the difficulty of obtaining seasonal or extreme weather on demand and the safety threats of adverse-weather field tests.Source: §10.2 · p. 32
Artificial tracks and chambers with rain- or snow-making machines provide controlled precipitation and lower-risk testing, but legislation limits level 4-or-higher AV tests in adverse weather.The facilities can approximate field conditions while controlling precipitation rates.Source: §10.2 · p. 32

11. Trends, Limitations & Future Research

This section examines future directions and limitations for adverse-weather autonomous driving, identifying sensor fusion, sophisticated learning methods, V2X/IoT, and alternative sensing approaches alongside gaps in dataset and tool coverage and barriers to adoption. It also notes perception degradation, 1550 nm LiDAR performance, and the role of technical competence and public trust in social acceptance.

11.1. Social significance

AV experience is associated with over 75% satisfaction and recognition and 60% more interest, while technical competence contributes to public trust in AVs.The section notes that legislation in multiple countries permits level 3 or 4 operations under conditions, while emergency human override or minimal-risk capability remains part of operating protocols [330].Source: §11.1 · p. 33

11.2.1. Toward advanced sensor fusion

Robust adverse-weather perception favors complementary sensor fusion, with future work needed to benchmark practical, cost-conscious combinations.The survey recommends selecting necessary sensors within economic and computational constraints and notes that LiDAR-dominant and thermal-camera-inclusive fusion have not been closely combined. FMCW LiDAR offers immunity to solar-light and cross-talk interference, while event cameras’ asynchronous output requires additional algorithms.Source: §11.2.1 · pp. 33–34

11.2.2. Toward sophisticated networks

In a nighttime raindrop-blurred-lens scenario, NVIDIA DRIVE active learning achieved over 3 times better accuracy and avoided false-positive detection.The approach selects unrecognized frames from unlabeled data for human annotation and adds them to the training pool; transfer and federated learning remain potential methods to explore.Source: §11.2.2 · p. 34

11.2.3. Toward V2X and IoT

V2V can relay weather and road-surface assessments and support warnings and rerouting; nearly 90% of test drivers found its front-collision warning and rerouting useful [277].V2I can provide weather alerts and clear-weather scene information to aid navigation, while V2X-aided perception and planning remain limited by connected-vehicle and infrastructure coverage. Combining V2V and V2I and processing road-weather hazards at the edge are proposed approaches; high-volume data transmission remains a constraint.Source: §11.2.3 · pp. 35–36

11.3. Limitations

Perception degradation is the main limitation of ADS sensors in adverse weather, while temperature and humidity endurance remain poorly analyzed.The survey summarizes sensor strengths and weaknesses across adverse conditions in a radar chart.Source: §11.3 · p. 36

11.3.1. 1550 nm LiDAR

In rain at 25 mm/hr, 905 nm LiDAR reaches twice as far as 1550 nm; in fog below 500 m visibility, it measures 60% farther.1550 nm has approximately 97% worse snow reflectance than 905 nm. The survey says the water-related attenuation and scarcity of adverse-weather field tests temper 1550 nm’s promise despite its strong-light filtering and interference resistance.Source: §11.3.1 · p. 37

11.3.2. Dataset and tool support

Current datasets adequately cover rain but have limited fog and snow support, while strong-light and contamination data are scarcer still.Fog data typically come from test facilities or simulators; snow support varies by falling versus accumulated snow, and artificial snow facilities are limited.Source: §11.3.2 · p. 37

11.4.1. Other LiDAR types

SWIR LiDAR can use material-specific spectral signatures for real-time multispectral detection and weather classification, but eye safety and adverse-weather field performance remain unclear.The passage notes uncertainty for super-continuum lasers operating at 728-810 nm and says few solid-state LiDARs are commercially ready for AVs or tested in adverse weather; newer spectrum-scan technology is anticipated.Source: §11.4.1 · p. 38

11.4.2. Other camera types

HDR cameras combine three exposures to capture dark and glare details, but moving objects can blur and video use may require extra time or reduced frame rate or resolution.A dual-gain CMOS sensor can avoid sacrificing frame rate or resolution, while ADS use may require additional image-processing algorithms.Source: §11.4.2 · p. 38
Hyperspectral cameras can identify material signatures across more than 100 wavelengths and are described as resistant to light-related conditions and seemingly unaffected by fog.Deployment is constrained by costs of $20,000-$100,000 USD and the number of devices needed for 360° coverage; the HSI-Drive dataset uses one 25-band camera and includes varied light, rainy/wet, and foggy conditions.Source: §11.4.2 · p. 38

11.4.3. Aerial view

Aerial LiDAR could extend AV perception beyond ground-level sight and support earlier risk avoidance, but real-time UAV-to-AV communication and incomplete coverage of tunnels and elevated roads hinder deployment.Its current use is mainly stationary tasks such as structural inspection.Source: §11.4.3 · p. 38

12. Conclusion

The survey synthesizes adverse weather effects on ADS sensors and AV functions, reviews hardware and perception solutions, and identifies research trends, future sensor candidates, and limitations. It finds progress in wet-weather perception but continuing gaps in snow, strong-light, and contamination research.

12. Conclusion

The survey covers common adverse weather, sensor effects, hardware and mechanical solutions, perception enhancement, auxiliary support, and weather-enabled datasets, simulators, and facilities.It also reviews classification, assessment, control and planning, and methods such as denoising.Source: §12 · p. 39
Robust sensor fusion, sophisticated networks and computer vision, and greater legislative and public support are identified as research tendencies; future sensor candidates include FMCW LiDAR, HDR cameras, and hyperspectral cameras.The survey also explains limitations from scarce relevant datasets and difficulties with 1550 nm LiDAR.Source: §12 · p. 39
Performance in common wet weather has progressed, with rain and fog improving through computer vision advances, although LiDAR still has room for improvement.Thermal cameras and radar are being deployed in sensor-fusion modalities to reinforce perception in adverse weather.Source: §12 · p. 39
Snow perception remains at the dataset-expansion and enhancement stage, motivating future point-cloud processing research in extreme snow under controlled or open-road interaction scenarios.Strong light and contamination also remain comparatively under-researched, with few solutions.Source: §12 · p. 39

mer, F. Dressler, Openc2x—an open source experimental and · all-weather capabilities. URL: https://www.laserfocusworld. com/lasers-sources/article/14035383/frequencymodulated-

This excerpt records a Laser Focus World URL concerning continuous-wave LiDAR’s all-weather capabilities and gives an access date of 14 June 2021.

all-weather capabilities. URL: https://www.laserfocusworld. com/lasers-sources/article/14035383/frequencymodulated-

The reference lists a Laser Focus World URL about continuous-wave LiDAR’s all-weather capabilities, accessed 14 June 2021.Source: all-weather capabilities. URL:… · p. 48