The correct estimation of the head pose is a problem of the great importancefor many applications. For instance, it is an enabling technology in automotivefor driver attention monitoring. In this paper, we tackle the pose estimationproblem through a deep learning network working in regression manner.Traditional methods usually rely on visual facial features, such as faciallandmarks or nose tip position. In contrast, we exploit a Convolutional NeuralNetwork (CNN) to perform head pose estimation directly from depth data. Weexploit a Siamese architecture and we propose a novel loss function to improvethe learning of the regression network layer. The system has been tested on twopublic datasets, Biwi Kinect Head Pose and ICT-3DHP database. The reportedresults demonstrate the improvement in accuracy with respect to currentstate-of-the-art approaches and the real time capabilities of the overallframework.
展开▼
机译：对于许多应用而言，正确估计头部姿势是非常重要的问题。例如，它是汽车中用于驾驶员注意力监控的一项启用技术。在本文中，我们通过以回归方式工作的深度学习网络来解决姿势估计问题。传统方法通常依赖于视觉面部特征，例如面部标志或鼻尖位置。相比之下，我们利用卷积神经网络（CNN）直接根据深度数据执行头部姿势估计。我们开发了一个暹罗体系结构，并提出了一种新颖的损失函数来改善回归网络层的学习。该系统已经在Biwi Kinect Head Pose和ICT-3DHP数据库这两个公共数据集上进行了测试。报告的结果表明，相对于当前最先进的方法和整体框架的实时功能，准确性有所提高。
展开▼
