Label correction in estimation method of winter road surface condition by deep learning based on highway data for adapting national roads

Abstract

In this paper, we propose a method for automatically correcting the results of a winter road surface condition estimation model using in-vehicle cameras. Typically, the estimation accuracy of road surface condition models decreases when applied to surfaces other than those used for training. However, collecting data and training the model for all road surfaces we wish to estimate is not feasible due to cost constraints. The proposed method mitigates this accuracy degradation without requiring model retraining. Specifically, when the confidence level of the predicted road surface is low, the method applies the k-nearest neighbors (k-NN) algorithm to the feature vectors of other road surface conditions within the model to relabel the prediction. The road surface conditions targeted for estimation are the six types commonly used in road management: Dry, Slightly wet, Wet, Slushy, Icy and Snowy. The proposed method was validated through experiments using real vehicle data collected on general roads, applying a model trained on winter highway data, and its effectiveness was confirmed.