Detecting Common Unsafe Gestures of Grabbing Wheel and Distracting Phone Talk by Appropriate Convolutional Neural Network

This study proposes a deep learning-based algorithm for identifying two most commonly happened gestures of a driver which are (1) grabbing the wheel with one hand and doing something with another hand, (2) talking on the phone while driving. These two gestures can certainly distract the driver from the attention of possible accidental events. The method involves dividing video into image frames, extracting gesture features, and training a neural network. This study compared and evaluated different neural models from Driver Action Expression (DAE) data set experimentally collected through a road driving experiment focusing on driver behavior. The problem of imbalanced data in the data set is also considered. The Xception model showed superior performance compared to other models, leading to its selection as the primary model for further research.