Detectable Object-Sizes Range Estimation Based Multi-Task Cascaded Convolutional Neural Networks in the Vehicle Environment

Abstract

In many studies regarding driver monitoring system, they directly employed face detection algorithms for general-purpose in an unconstrained environment. These algorithms are generally not suitable for limited resources of vehicles. Unlike the general unconstrained environment, the range of detectable face sizes and locations can be estimated in the vehicle environment. In this paper, we propose the Detectable Object-sizes Range Estimation algorithm (DORE) to estimates the range of detectable face sizes through specific information in the vehicle environment. The DORE algorithm makes images, in which a face is likely to be detected in the in-vehicle environment, to be fed into a face detection algorithm, such as Multi-task cascaded Convolutional Neural Networks (MTCNN) which stably detect faces rather than others. Our experiment shows that DORE applied MTCNN not only had the same performance as MTCNN in terms of accuracy but also had relatively low processing time in the vehicle environment.