Lateral Road-mark Reconstruction Using Neural Network for Safe Autonomous Driving in Snow-wet Environments

This paper highlights the importance of utilizing artificial intelligence to safely localize autonomous vehicles in hard weather and environmental conditions. An edge profile is created to encode the lateral road-marks as a series of peaks. A neural network is utilized to explore different relationships between the peaks among a set of map edge profiles. Based on the generalization capability of the trained network, the LIDAR profiles are reconstructed during the autonomous driving to regularly reform the peaks in terms of distribution and amplitude patterns. A reconstructed profile is then matched to the corresponding map profile to estimate the lateral vehicle position accurately. The proposed framework has been tested in very hard situations of snow and grass accumulation at the roadside, snow lines inside lanes and, on wet and old road surfaces. The experimental results have verified the reliability and safety of reconstructing LIDAR profiles to achieve precise lateral localization against the changes of road, weather and environmental conditions.