Non‐instinct detection of cellphone usage from lane‐keeping performance based on eXtreme gradient boosting and optimal sliding windows

Abstract

Abstract Driving distraction caused by cellphone usage has become a common safety threat. As distraction detection methods based on driver's position or eye movement may raise privacy issues, a promising way is to analyze the vehicle's lane‐keeping performance. This paper proposed a detection algorithm based on eXtreme gradient boosting (XGBoost), to develop a real‐time driving distraction detection based on lane‐keeping performance. The algorithm includes knowledge‐based volatility feature extraction and feature selection by recursive feature elimination (RFE). To obtain dynamic patterns of lane‐keeping performance affected by different types of cellphone usage, browsing a short message, browsing a long message, and answering a phone call, a driving simulator experiment was conducted on 28 drivers. Results showed that the proposed XGBoost‐RFE method is reliable and promising to predict phone usage with 80% accuracy. The results also evoke the fact that sliding window size, which is about 80% of subtask duration, can be appropriate for real‐time detection of multiple cellphone usages. For overlap percentages, 67% of sliding window size can balance the efficiency and continuity of data in adjacent sliding windows. The paper's potential application includes the design of a real‐time driving distraction detection system.