视觉协同的违规驾驶行为分析方法

PDF下载HTML

Title:

A visual collaborative analysis method for detecting illegal driving behavior

作者:

高尚兵^1,2, 黄子赫¹, 耿璇¹, 臧晨¹, 沈晓坤¹

1. 淮阴工学院 计算机与软件工程学院，江苏 淮安 223001;
2. 淮阴工学院 江苏省物联网移动互联技术工程实验室，江苏 淮安 223001

Author(s):

GAO Shangbing^1,2, HUANG Zihe¹, GENG Xuan¹, ZANG Chen¹, Shen Xiaokun¹

1. College of Computer and Software Engineering, Huaiyin Institute of Technology, Huaian 223001, China;
2. Laboratory for Internet of Things and Mobile Internet Technology of Jiangsu Province, Huaiyin Institute of Technology, Huaian 223001, China

关键词:

驾驶行为识别; 模型剪枝; 目标检测; 姿态估计; 协同检测; 模型优化; 深度学习; 卷积神经网络

Keywords:

driving behavior recognition; model pruning; target detection; attitude estimation; collaborative detection; model optimization; deep learning; convolutional neural network

分类号:

TP391.4

DOI:

10.11992/tis.202101024

摘要:

本文针对危险驾驶识别中主流行为检测算法可靠性差的问题，提出了一种快速、可靠的视觉协同分析方法。对手机、水杯、香烟等敏感物体进行目标检测，提出的LW(low weight)-Yolov4(You only look once v4)通过去除CSPDarknet53(cross stage partial Darknet53)卷积层中不重要的要素通道提升了检测速度，并L₁正则化产生稀疏权值矩阵，添加到BN(batch normalization)层的梯度中，实现优化网络模型的目的；提出姿态检测算法对驾驶员指关节关键点进行检测，经过仿射逆变换得到原始帧中的坐标；通过视觉协同分析对比敏感物品的检测框位置与驾驶员手部坐标是否重合，判定驾驶员是否出现违规驾驶行为及类别。实验结果表明，该方法在识别精度与检测速度方面均优于主流的算法，能够满足实时性和可靠性的检测要求。

Abstract:

This study proposes a fast and reliable visual collaborative analysis method to improve the reliability of mainstream behavior detection algorithms in dangerous driving recognition. First, the algorithm performs target detection on sensitive objects such as mobile phones, water cups, and cigarettes. The proposed low weight-Yolov4 algorithm improves the detection speed by removing unimportant element channels in the cross-stage partial Darknet53 convolutional layer and regularizes L₁ to generate a sparse weight matrix. Besides, the obtained matrix is added to the gradient of the batch normalization layer to optimize the network model. Then, an attitude detection algorithm is used to detect key points of the driver’s knuckles, and the coordinates in the original frame are obtained through the affine inverse transformation. Finally, the driver’s illegal driving behavior and its category are determined through visual collaborative analysis and comparison of the position of the detection frame of sensitive objects and coordinates of the driver’s hands. Experimental results show that the recognition accuracy and detection speed of the proposed method are better than those of mainstream algorithms, which can meet the detection requirements of real-time and reliability.