An Improved Single Shot Multibox for Video-Rate Head Pose Prediction

Abstract

Head pose estimation plays a crucial role in attention detection, behavior analysis, human-computer interaction, and eye tracking, etc. The existing landmark-to-pose methods require two steps, not only to detect the key points of the human face, but also to solve the 2D to 3D mapping problem usually by the average head model. We propose an efficient and robust method to monitor the driver's attention. Firstly, the prevailing object detection algorithm SSD which has inherent capabilities of simultaneous classify and regress, is used to create a lightweight network, which avoids the shortcomings of high coupling and time-consuming of the existing methods. Then, single-scale anchors, which have a less computational cost than multi-scale anchors, are adopted for vehicle environments where the ambient light changes dramatically. Finally, by binning continuous angles into specific classes, the 3D angle regression problem is converted into angle classification and face box regression, and our model directly outputs Euler angles (Yaw, Pitch, and Roll) without detecting face landmarks. Experiments on YawDD result that our approach can efficiently perform detection tasks and estimation tasks under the actual driving environment of various luminosity. The mean average errors of prediction in AFLW2000 and 300W-LP are 6.01° and 2.38°, which demonstrates the accuracy of the proposed algorithm.