Real-Time Head Orientation from a Monocular Camera Using Deep Neural Network

We propose an efficient and accurate head orientation estimation algorithm using a monocular camera. Our approach is leveraged by deep neural network and we exploit the architecture in a data regression manner to learn the mapping function between visual appearance and three dimensional head orientation angles. Therefore, in contrast to classification based approaches, our system outputs continuous head orientation. The algorithm uses convolutional filters trained with a large number of augmented head appearances, thus it is user independent and covers large pose variations. Our key observation is that an input image having 32 x 32 resolution is enough to achieve about 3 degrees of mean square error, which can be used for efficient head orientation applications. Therefore, our architecture takes only 1ms on roughly localized head positions with the aid of GPU. We also propose particle filter based post-processing to enhance stability of the estimation further in video sequences. We compare the performance with the state-of-the-art algorithm which utilizes depth sensor and we validate our head orientation estimator on Internet photos and video.