DarNet: A deep learning solution for distracted driving detection

Abstract

Distracted driving is known to be the leading cause of motor vehicle accidents. With the increase in the number of IoT devices available within vehicles, there exists an abundance of data for monitoring driver behavior. However, designing a system around this goal presents two key challenges - how to concurrently collect data spanning multiple IoT devices, and how to jointly analyze this multimodal input. To that end, we present a unified data collection and analysis framework, DarNet, capable of detecting and classifying distracted driving behavior. DarNet consists of two primary components: A data collection system and an analytics engine. Our system takes advantage of advances in machine learning (ML) to classify driving behavior based on input sensor data. In our system implementation, we collect image data from an inward facing camera, and Inertial Measurement Unit (IMU) data from a mobile device, both located within the vehicle. Using deep learning techniques, we show that DarNet achieves a Top-1 classification percentage of 87.02% on our collected dataset, significantly outperforming our baseline model of 73.88%. Additionally, we address the privacy concerns associated with collecting image data by presenting an alternative framework designed to operate on down-sampled data which produces a Top-1 classification percentage of 80.00%.