激光与光电子学进展, 2020, 57 (6): 061019, 网络出版: 2020-03-06
基于导师-学生网络的驾驶行为识别方法 下载: 949次
Driving Behavior Recognition Method Based on Tutor-Student Network
图像处理 深度学习 驾驶行为识别 卷积神经网络 弱定位 动作分类 导师-学生网络 image processing deep learning driving behavior recognition convolutional neural network weak location action classification tutor-student network
摘要
提出了一种基于导师-学生网络的驾驶行为识别模型。考虑到驾驶动作是在局部区域发生的,将驾驶行为识别任务拆分成动作定位和动作分类两个子任务。针对动作定位任务,设计了网络层数较浅和接收高分辨率图像输入的导师网络,导师网络通过特征图的响应对动作区域进行弱定位;在动作定位基础上,针对动作分类任务,设计了网络层数较深的接收低分辨率动作区域图像输入的学生网络,学生网络根据深层网络提取的高层次语义特征实现高准确率分类。实验结果证明,导师-学生网络模型能带来较高的识别准确率,具有强稳健性。
Abstract
This paper presents a driving behavior recognition model based on tutor-student network. Considering that driving behavior occurs in a local area, this paper divides the task of driving behavior recognition into two sub-tasks: action location and action classification. Aiming at the task of action location, a tutor network with shallow network layer receiving high-resolution image input is designed. The tutor network weakens the action area according to the response of feature map. On the basis of action location and action classification task, a student network with deeper network layer is designed to receive the input of low-resolution action area image. High-level semantic features which are extracted from student network are used to achieve high accuracy classification. Experimental results show that the tutor-student network model can bring high recognition accuracy and strong robustness.
褚晶辉, 张姗, 汤文豪, 吕卫. 基于导师-学生网络的驾驶行为识别方法[J]. 激光与光电子学进展, 2020, 57(6): 061019. Jinghui Chu, Shan Zhang, Wenhao Tang, Wei Lü. Driving Behavior Recognition Method Based on Tutor-Student Network[J]. Laser & Optoelectronics Progress, 2020, 57(6): 061019.