Driver's gaze zone estimation by transfer learning
Driver's gaze zone estimation by transfer learning
复制标题
通过迁移学习估计驾驶员的注视区域
DOI：
10.1109/icce.2018.8326308
复制
发表时间：
2018
期刊：
影响因子：
--
通讯作者：
Soon Kwon
中科院分区：
文献类型：
--
作者：
Iman Rahmansyah Tayibnapis;Min;Soon Kwon
驾驶员视线区域的估计对于高级驾驶员辅助系统（ADAS）的实现具有重要的意义。注视估计可以监视驾驶员焦点并且使用增强现实平视显示器（AR-HUD）间接地控制挡风玻璃上的用户界面/用户体验（UI/UX）。然而，为了训练视线区域估计器作为分类任务，有人付出了巨大的成本来收集大量的注释数据集。为了减少劳动力工作，我们使用了一种使用预训练CNN模型的迁移学习方法，通过在具有大而可靠的数据集的移动的设备上进行回归来将注视估计任务投影到新的分类任务中，以克服缺乏用于注视区估计的注释数据集的问题。我们测试了我们自己的建筑模拟测试床所提出的方法。结果显示，对于估计车内10个凝视区，验证准确度约为99.01%，测试准确度约为60.25%。
Estimating driver's gaze zone has very important role to support advanced driver assistant system (ADAS). The gaze estimation can monitor the driver focus and indirectly control the user interface/user experience (UI/UX) on a windshield using augmented reality-head up display (AR-HUD). However, to train gaze zone estimator as a classification task, someone pays huge costs to gather a large amount of annotated dataset. To reduce the labor work, we used a transfer-learning method using pre-trained CNN model to project the gaze estimation task by regression on mobile devices that have large and reliable dataset into new classification task to overcome lack of annotated dataset for gaze zone estimation. We tested the proposed method to our own building simulation test bed. The result is shown in validation accuracy around 99.01 % and test accuracy with unseen driver around 60.25 % for estimating 10 gaze zones in-vehicle.