A machine learning model pipeline for detecting wet pavement condition from live scenes of traffic cameras

Abstract

Highway safety is largely influenced by weather conditions that have become increasingly volatile due to the climate change. It well known that wet pavement significantly reduces surface friction, leading to inflated collision risk. Thus, timely knowledge of the road surface condition is critical for safe driving. In this paper, a novel machine learning model pipeline is proposed to detect the wetness of pavement based on live images of highway scenes captured by publicly accessible traffic cameras. To simplify the learning task, we finetuned the state-of-the-art instance segmentation baseline models to extract background instance targets, including pavement, sky, and vegetation, which are common in highway scenes. Then, the color mixture attributes in HSV (hue, saturation and value) of each segmented instance were extracted and used as visual cues for inferring pavement condition. Finally, gradient boosting ensemble classifiers are constructed and trained using the HSV features to predict the wetness of pavement. For the segmentation task, we leveraged Detectron2 baseline models (Mask R-CNN) and evaluated three backbone networks: R50-FPN, R101-FPN, and X101-FPN. For the classification task, two most popular gradient boosting algorithms (XGBoost and CatBoost) were evaluated together with a classic logistic model. Based on experiments with our custom dataset, the best performance (F1 score: 0.927, AUC: 0.975) was achieved by the R101-FPN backbone coupled with the CatBoost classifier.

Keywords

• Machine learning
• Image segmentation
• Convolutional neural networks
• Gradient boosting
• CatBoost
• XGBoost