Dept. of Robotics Science and Technology,
Chubu University

Deep Learning Conference

Body Posture and Face Orientation Estimation by Convolutional Network with Heterogeneous Learning

Author
K. Okuno, T. Yamashita, H. Fukui, S. Noridomi, K. Arata, Y. Yamauchi and H. Fujiyoshi
Publication
International Workshop on Advanced Image Technology, 2018

Download: PDF (English)

Autonomous driving system switches over to a manual driving mode by human when the system is not able to drive itself. The system has to constantly monitor whether the driver can drive the vehicle by the driver’s posture and face orientation. Conventional methods for estimating posture and face orientation perform feature extraction and recognition for each task, and thus require an appreciable amount of processing time. In this paper, we propose a method that performs multiple tasks by Deep Convolutional Neural Network (DCNN) with heterogeneous learning, by sharing the feature extraction process. The body posture and face orientation estimation can be performed simultaneously. In evaluation, we have achieved a high accuracy of 98% in body posture estimation, and 91% in face orientation estimation. The processing time for a single image has been 2.6 ms when the we employ a GPU, and 34.1 ms in CPU. We confirm that proposed method can perform body posture and face orientation estimation in real time.

Previous Next