Driver activity recognition using deep learning based on multi-step batch size up

Abstract

The increasing popularity of electric motorbikes in Indonesia, while promoting sustainable mobility, also raises concerns regarding traffic safety. Given the high incidence of motorcycle-related accidents, there is a critical need for systems capable of monitoring and recognizing driver behavior. This study proposes a driver activity recognition system for electric motorbikes, utilizing an event data recorder (EDR) to capture seven key sensor signals: three-axis acceleration, voltage, current, power, and speed. A custom dataset was constructed using data collected from 10 subjects, each performing five driving activities including forward drive, brake, stop, left turn, and right turn for over three-minute intervals per activity. The classification model is based on a long short-term memory (LSTM) neural network. To optimize training efficiency, a multi-step batch size up (MSBU) strategy was introduced, which accelerates training time by 1.84× compared to a fixed batch size of 32. The best performance was achieved using a segment length of 75 time-steps, yielding an accuracy and macro F1-score of 0.9873. These results demonstrate the effectiveness of the proposed system for real-time driver behavior monitoring and activity recognition in electric motorbike applications.