汽车工程 ›› 2021, Vol. 43 ›› Issue (8): 1203-1209.doi: 10.19562/j.chinasae.qcgc.2021.08.011

• • 上一篇    下一篇

基于改进长短时记忆网络的驾驶行为检测方法研究

施冬梅1(),肖锋2   

  1. 1.苏州信息职业技术学院计算机科学与技术系,苏州 215200
    2.西安工业大学计算机科学与工程学院,西安 710021
  • 收稿日期:2021-03-02 修回日期:2021-05-30 出版日期:2021-08-25 发布日期:2021-08-20
  • 通讯作者: 施冬梅 E-mail:sdm1976@126.com
  • 基金资助:
    国家自然科学基金(61572394);陕西省科技计划项目(2020GY-066);江苏省自然科学基金(BK20191225);2020年苏州高职高专第二批产学研合作基地项目(2020-5)

Study on Driving Behavior Detection Method Based on Improved Long and Short⁃term Memory Network

Dongmei Shi1(),Feng Xiao2   

  1. 1.Department of Computer Science and Technology,Suzhou College of Information Technology,Suzhou 215200
    2.School of Computer Science and Engineering,Xi’an Technological University,Xi’an 710021
  • Received:2021-03-02 Revised:2021-05-30 Online:2021-08-25 Published:2021-08-20
  • Contact: Dongmei Shi E-mail:sdm1976@126.com

摘要:

疲劳驾驶和不安全驾驶行为是引起交通事故的主要原因,随着智能交通技术的发展,利用深度学习算法进行驾驶行为检测已成为研究的热点之一。在卷积神经网络和长短时记忆神经网络的基础上,结合注意力机制改进网络结构,提出一种混合双流卷积神经网络算法,空间流通道采用卷积神经网络提取视频图像的空间特征值,以空间金字塔池化代替均值池化,统一了特征图的尺度变换,时间流通道采用SSD算法计算视频序列相邻两帧光流图像,用于人眼等脸部小目标的检测,再进行图像特征融合与分类,在LFW数据集和自建数据集中进行了实验,结果表明本方法的人脸识别和疲劳驾驶的检测准确率分别高于其他方法1.36和2.58个百分点以上。

关键词: 安全驾驶, 卷积神经网络, 长短时记忆, 单步检测, 人脸识别, 疲劳驾驶检测

Abstract:

Fatigue driving and unsafe driving behavior are the main causes of traffic accidents. With the progress of intelligent transportation technology, using deep learning algorithm to detect driving behavior has become one of the hotspots of research. On the basis of convolution neural network (CNN) and long?term memory neural network, a hybrid dual stream convolution neural network algorithm is proposed by combining attention mechanism to improve network structure. In spatial flow channel, CNN is used to extract the spatial feature values of video image and the traditional mean pooling is replaced by spatial pyramid pooling is used to replace mean pooling, with the transformation of feature map unified. In time stream channel, single shot detection algorithm is adopted to calculate two adjacent frames of optical flow images of video sequence for detecting small facial targets such as human eyes. Then the fusion and classification of image features are carried out. Finally, experiments are performed on LFW dataset and self?built dataset. The results show that with the method adopted the accuracy of face recognition and fatigue driving detection is 1.36 and 2.5 percentage points higher than other methods respectively.

Key words: safe driving, CNN, LSTM, SSD, face recognition, fatigue driving detection