基于滑动窗特征融合的深信度网络驾驶行为识别

Driving behavior recognition algorithm for deep belief network based on sliding window feature fusion

针对动态突变行为原始信息量较少造成的行为不易区分以及浅层结构分类算法分类正确率较低的问题,提出一种基于滑动窗特征融合的深信度网络驾驶行为识别算法。采用从手机传感器中获取的三轴加速度数据进行预处理后,利用端点检测算法确定行为切换点,通过滑动窗实时提取时间序列信息并计算序列片段的时频域特征,选取有效特征后,融合原始行为信息与特征建立完整时间序列段作为受限玻尔兹曼机的输入,优化预设参数的多隐层受限玻尔兹曼机对输入端信息的特征进行提取,最终通过深信度网络(deep belief network,DBN)实现驾驶行为的识别。实验结果表明,改进的滑动窗特征融合的深信度网络驾驶行为识别算法整体识别率为98.4%,能有效进行驾驶行为的识别。

This paper proposed a driving behavior recognition algorithm based on feature fusion of acceleration data for deep belief network to solve the problem of the original small amount of information for dynamic mutation behavior,difficult to distinguish the behavior and took the major classification algorithms into consideration.It obtained the three-axis acceleration data from the mobile phone and after pretreatment utilizing endpoint detection algorithm to determine the switching point,the time-frequency domain feature calculated when time-series information extracted through a sliding window in real time,fusing original signals with feature as restricted Boltzmann machine’s input,optimized the multiple hidden layers restricted Boltzmann machine for input’s feature information extraction,and then through the DBN to achieve recognition.Driving behavior results show that the improved driving behavior recognition accuracy is 98.4%,which can effectively identify driving behavior.