Snow Detection using In-Vehicle Video Camera with Texture-Based Image Features Utilizing K-Nearest Neighbor, Support Vector Machine, and Random Forest

Abstract

Snowfall negatively affects pavement and visibility conditions, making it one of the major causes of motor vehicle crashes in winter weather. Therefore, providing drivers with real-time roadway weather information during adverse weather is crucial for safe driving. Although road weather stations can provide weather information, these stations are expensive and often do not represent real-time trajectory-level weather information. The main motivation of this study was to develop an affordable in-vehicle snow detection system which can provide trajectory-level weather information in real time. The system utilized SHRP2 Naturalistic Driving Study video data and was based on machine learning techniques. To train the snow detection models, two texture-based image features including gray level co-occurrence matrix (GLCM) and local binary pattern (LBP), and three classification algorithms: support vector machine (SVM), k-nearest neighbor (K-NN), and random forest (RF) were used. The analysis was done on an image dataset consisting of three weather conditions: clear, light snow, and heavy snow. While the highest overall prediction accuracy of the models based on the GLCM features was found to be around 86%, the models considering the LBP based features provided a much higher prediction accuracy of 96%. The snow detection system proposed in this study is cost effective, does not require a lot of technical support, and only needs a single video camera. With the advances in smartphone cameras, simple mobile apps with proper data connectivity can effectively be used to detect roadway weather conditions in real time with reasonable accuracy.