A public domain dataset to recognize driver entry into and exit from a car using smartphone sensors

Human activity dataset collected via smartphone sensors have been popular in identifying activities of daily living and creating situational awareness. Various classical classifiers and more recently deep learning techniques have been used over such datasets to deduce human behaviour including driver behaviour, rash driving, accident detection and road conditions. However, no public smartphone dataset exists which can help recognise driver entry and exit from a car. Such a dataset is important for early detection of driving activity and can help to activate accident detection mechanism as soon as a driver boards a car. This study presents a public dataset which records four activities viz. a driver getting IN/OUT and SITTING in/STANDING out of a car. The dataset is labeled and presented in two forms, both with and without applying feature engineering. Performance of different base and meta-level classifiers like logistic regression, linear SVC, SVM, decision tree, random forest, k-NN and gradient boosting DT is evaluated. With hyper parameter tuning the parameters of the best estimator are recorded. Classification accuracy from 86.97 to 95.33% is observed for different classifiers on the dataset.