Dilated Convolutional Neural Network for Predicting Driver's Activity

摘要

Anticipation and prediction play a pivotal role in human-like perception and memory. Anticipating human activities is an essential capability for any natural and seamless human-robot interaction scenario. To effectively understand and reason about human activities, a smart system needs to have the ability to process sequential/temporal observations that are normally noisy, high dimensional, have long temporal dependencies and have a high frequency (e.g, videos). In this paper, we propose a novel deep learning model architecture to classify driver's actions and activities in real-world scenarios of driving a car in different conditions. Sensory data comes from a variety of sources including a driver facing camera (inside camera), a road facing camera (outside camera), GPS and other car related sensors. The proposed model is flexible and easy to use since it is not relied on external methods to extract key-points from video frames. It uses convolution and max-pooling pairs to understand spatial relationships within video frames and incorporates dilated deep convolutional structures to capture long temporal dependencies, process and predict driver's activities. We show the results to compete with the state of the art in this domain.