电子测量与仪器学报

程鸣,严运兵.驾驶员视线区域自适应半监督学习标定方法[J].电子测量与仪器学报,2023,37(12):136-142

驾驶员视线区域自适应半监督学习标定方法

Adaptive semi supervised learning calibration method for driver’s line of sight region

DOI：

英文关键词:line of sight area semi supervised learning small sample data calibration techniques adaptive adjustment

基金项目:国家自然科学基金(51975428)项目资助

作者	单位
程鸣	1.武汉科技大学汽车与交通工程学院
严运兵	1.武汉科技大学汽车与交通工程学院

Author	Institution
Cheng Ming	1.School of Automotive and Traffic Engineering, Wuhan University of Science and Technology
Yan Yunbing	1.School of Automotive and Traffic Engineering, Wuhan University of Science and Technology

摘要点击次数: 206

全文下载次数: 433

中文摘要:

目前监测驾驶员视线区域的算法通常采用深度学习端模型直接对图像特征分类,此方法依赖固定座舱视角下采集的驾驶员视线区域数据,但由于驾驶员外形特征差异、坐姿习惯差异和摄像头安装位置差异的影响,难以获取大量且全面的数据,导致分类精度降低的问题,如何仅采用小样本数据集提升视线区域识别精度成为难题。本文将基于半监督学习理论设计一种自适应的视线区域标定方法。首先采用 L2CS 模型回归小样本数据中驾驶员视线角度二维向量,再通过统计分析挖掘驾驶员视线角度和视线区域映射的泛化先验知识,利用该知识进行视线区域标定,剔除非待检区域的无效视线落点,并以滑动窗口方式完成针对驾驶员个人的视线区域精细化分类。经试验证明,该方法解决了端模型数据跨域能力低下的问题,准确率和召回率分别提升 22. 4%和 10. 3%,且标定结果具有自适应修复能力。

英文摘要:

At present, algorithms for monitoring the driver’s line of sight area usually use deep learning models to directly classify image features. This method relies on the driver’ s line of sight area data collected from a fixed cockpit perspective. However, due to differences in driver appearance, sitting habits, and camera installation positions, it is difficult to obtain a large amount of comprehensive data, resulting in a decrease in classification accuracy. How to improve the accuracy of line of sight recognition using only small sample datasets has become a challenge. This article will design an adaptive line of sight region calibration method based on semi supervised learning theory. Firstly, the L2CS model is used to regress the two-dimensional vector of driver’ s line of sight angle in small sample data. Then, statistical analysis is used to mine the generalization prior knowledge of driver’ s line of sight angle and line of sight area mapping. This knowledge is used for line of sight area calibration, removing invalid line of sight landing points in non-inspection areas, and completing fine classification of driver’s personal line of sight area in a sliding window manner. Through experiments, it has been proven that this method solves the problem of low cross domain capability of end model data, improving accuracy and recall by 22. 4% and 10. 3% respectively, and the calibration results have adaptive adjustment ability.

查看全文查看/发表评论下载PDF阅读器