FRNet: DCNN for Real-Time Distracted Driving Detection Toward Embedded Deployment
Cong Duan, Yipeng Gong, Jiacai Liao, Minghai Zhang, Libo Cao
Abstract
Real-time running deep convolutional neural networks on embedded electronics is one recent focus for distracted driving detection. In this work, we proposed <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\textit {FRNet}$ </tex-math></inline-formula> , which is a unique, efficient, and real-time architecture. The <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\textit {FRNet}$ </tex-math></inline-formula> converts spatially distributed features into depth distribution by a feature reorganization block. This block compresses the volume of backbone, reduces the memory read/write volumes along with multiply-accumulate operations, and extracts key features faster. In addition, a ultra-lightweight backbone was designed, with an atypical reshape strategy. This atypical strategy designed based on pixel-level analysis, for compensating the accuracy decline along with feature reorganization. The proposed <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\textit {FRNet}$ </tex-math></inline-formula>  offered excellent real-time performance on low power embedded platforms, and has a competitive accuracy with previous state-of-the-art models. It achieved 97.55% accuracy on the <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\textit {SFD+AUCDD-V1}$ </tex-math></inline-formula> , and 99.86% on <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$\textit {3MDAD}$ </tex-math></inline-formula> . On an automotive-grade embedded demo board, it costs 16.93 ms per frame and achieved 59 FPS. As of today, this is the fastest record for end-to-end distraction detection. Experiments revealed the real-time and accuracy are best balanced with FRNet. The model is publicly available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/congduan-HNU/FRNet</uri> .