A dynamic window-size based segmentation technique to detect driver entry and exit from a car
Amit Hirawat, Swapnesh Taterh, Tarun Kumar Sharma · Journal of King Saud University - Computer and Information Sciences · 2021
In-built hardware sensors of a smartphone have emerged as a useful sensor platform for human activity recognition and context awareness. Most of available research in human activity recognition domain is expressed as a classification problem where a classifier’s performance is directly co-related upon the amount of inter-class variance which feature engineering can bring. Several feature extraction techniques have been used by existing studies, including dynamic and static activity-window sizes. In our study, we propose a dynamic non-overlapping window based approach which spans across the entire time-series data representing the activity (DyNOEA). The objective is to detect driver entry and exit from a car using smartphone sensors, while the smartphone is in the driver’s pocket. In comparison to earlier works, which used static and overlapping window-size for activity recognition, the proposed feature extraction technique is able to classify activities with extremely high accuracy in perfectly naturalistic setup achieving up to 100% accuracy for classifiers like logistic regression, decision tree and k-NN.