Core Concepts
A convolutional neural network model can accurately predict dense pixelwise grip maps of the road area by fusing data from RGB cameras, thermal cameras, and LiDAR reflectance sensors.
Abstract
The authors propose a method to predict dense grip maps of the road surface ahead of an autonomous vehicle using a convolutional neural network (CNN) model. The model takes as input fused data from RGB cameras, thermal cameras, and LiDAR reflectance sensors, and is trained to predict pixelwise grip values using sparse ground truth measurements from an optical road weather sensor.
The key highlights of the work are:
Data Collection and Preprocessing:
The authors collected a 37-hour, 1538 km dataset of driving data in various weather conditions, including snow, rain, and ice.
They developed a method to precisely align and match the data from the different sensors (RGB, thermal, LiDAR) with the sparse ground truth grip measurements from the road weather sensor.
Model Architecture:
The authors used a Feature Pyramid Network (FPN) architecture with separate encoders for each input modality, fusing the features at different scales.
The model was trained to predict both the grip value and the surface layer thicknesses (water, ice, snow) as an auxiliary task.
Evaluation:
The model achieved good grip prediction accuracy, with root mean square errors (RMSE) of 0.0632 on the validation set and 0.0575 on the test set.
The qualitative results show the model can accurately capture the boundaries between different road surface conditions, such as clear tire tracks on snowy roads.
Fusing the RGB, thermal, and LiDAR data improved the grip prediction accuracy compared to using any single modality.
The authors conclude that the proposed method can effectively predict dense grip maps of the road area ahead of the vehicle, which could enable autonomous driving systems to better navigate in adverse weather conditions.
Stats
24% of weather-related vehicle crashes in the U.S. occur on snowy, slushy, or icy pavement.
15% of weather-related vehicle crashes happen during snowfall or sleet.
The dataset contains 237,067 samples collected over 37 hours (1538 km) in various weather conditions.
The validation set has 15,343 samples, the test set has 26,783 samples, and there are 3 additional test drives with 16,139 samples in total.
Quotes
"Slippery road weather conditions are prevalent in many regions and cause a regular risk for traffic."
"Besides low visibility, significant challenges posed by winter conditions are changes in road surface slipperiness."