基于BP神经网络的无校准驾驶员注视区域估计

发布时间：2018-07-03 00:33

本文选题：BP神经网络 + 区域分类器　；参考：《大连海事大学》2017年硕士论文

【摘要】：二十世纪以来,汽车的拥有量显著增长,随之而来的交通事故也发生的越来越频繁,人们对交通安全问题的关注与日俱增,世界各国的交通部正积极采取一些有效措施,减少交通事故的发生。因此,基于驾驶员的视线追踪系统便应运而生。但现有的系统一般局限于简单场景,且必须在前期做好校准工作的情况下进行的,而对于不加约束的人、无校准、光照变化等问题仍存在很大的研究空间,其实时性、精确性和鲁棒性与实际应用之间还存在很大距离。针对这些问题,本文提出基于BP神经网络的无校准驾驶员注视区域估计方法,将重点从头部姿态和瞳孔视线角度参数获取、BP神经网络注视区域估计算法,及实验对比评估三个部分进行研究,主要工作如下:首先,本文需要先获取头部姿态和视线方向的角度参数,在此过程中,针对驾驶员身体出现左右晃动,或者不同驾驶员身高不同,而发生相对于摄像机的左右偏移和上下偏移问题,提出一种基于几何关系的头部姿态校正算法。同时,本文通过建立3D眼球模型进行瞳孔视线方向估计。然后,构建了一个基于BP神经网络的无校准驾驶员注视区域估计系统。通过BP神经网络模型对驾驶员在驾驶过程中的头部姿态及视线角度参数进行训练并构建区域分类器,并通过该网络模型进行驾驶员无校准注视区域估计。最后,对本文方法进行评估。通过对比实验表明,本文提出的方法不仅能满足学术研究的要求,而且能实现在复杂环境下驾驶员的注视区域估计,满足了对实验的实时性,精确度和鲁棒性的要求,并为安全驾驶的辅助系统打下良好的基础。
[Abstract]:Since twentieth Century, the number of cars has increased significantly, and the traffic accidents are becoming more and more frequent. People pay more attention to traffic safety. The transportation department of the world is actively taking some effective measures to reduce the occurrence of traffic accidents. However, the existing systems are generally limited to simple scenes, and must be carried out in the early stage of calibration, but there is still a lot of research space for unconstrained people, no calibration, light change and other problems, and there is a great distance between the reality, the accuracy and the robustness. An uncalibrated area estimation method based on BP neural network, which focuses on the head attitude and the eye view angle parameters, the BP neural network fixation area estimation algorithm, and the experimental comparison and evaluation of three parts are studied. The main work is as follows: first, we need to obtain the angle reference of the head attitude and the direction of sight. In this process, a head attitude correction algorithm based on the geometric relationship is proposed for the driver body sloshing, or different driver's height, and the left and right offset and up and down migration of the camera. At the same time, the 3D eye model is established to estimate the eye direction of the pupil. Then, the construction of the eye direction is constructed. An uncalibrated area estimation system based on BP neural network is introduced. The BP neural network model is used to train the driver's head attitude and view angle parameters in the driving process and construct a regional classifier, and the driver's uncalibrated gaze area estimation is carried out through the network model. Finally, the method is carried out. A comparative experiment shows that the proposed method can not only meet the requirements of academic research, but also realize the estimation of the driver's gaze area under the complex environment, meet the requirements of the real-time, accuracy and robustness of the experiment, and lay a good foundation for the auxiliary system of safe driving.
【学位授予单位】：大连海事大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【参考文献】