An adaptive machine learning framework for multi-scenes road surface weather condition monitoring

Abstract

Timely road surface condition (RSC) monitoring and maintenance significantly influences road safety. The current RSC relies on fixed road surveillance cameras and in-vehicle cameras. However, the fixed camera demands higher precision, while the in-vehicle camera requires higher timeliness. To address these challenges, this paper introduces an adaptive machine learning framework for simultaneous road surface detection on both device types. Initially, a convolutional neural network-based differentiation module identifies image sources. Subsequently, an adaptive algorithm switching mechanism leads to the development of two algorithms improved upon the real-time object detection algorithms. At last, extensive experiments with datasets collected from Ontario, Canada and Iowa, U.S.A. validate the framework. Results show satisfactory classification accuracy, detection precision, and speed. Notably, the mean average precision, namely mean of the average precision for all categories (mAP), reaches 91.9% for fixed cameras and 90.6% for in-vehicle cameras, outperforming existing road surface snow detection models.