基于深度学习的安全带检测方法研究

发布时间：2018-08-18 16:56

【摘要】：近年来,我国机动车数量不断增加,交通事故及伤亡人数也是呈逐年上升趋势。安全带作为一种十分重要的被动保护措施,可有效的降低车辆在道路上行驶时因车辆碰撞或其他交通事故造成的伤亡率。我国交管部门及相关的法律法规严格要求车辆驾驶人员在车辆行驶过程中需佩带安全带。但在我国驾驶人员在行车过程中不系安全带的现象还是普遍存在,主要原因为驾驶员安全意识不强,存在多种躲避安全带提示系统的不规范行为。因此,研究机动车内驾驶人员是否佩戴安全带的检测方法对于提高驾驶员的遵守交通法规的意识具有十分重要的意义。本文提出了一种基于深度学习的安全带检测方法,尝试将近年来在图像识别方面有较好应用的深度学习方法来提高安全带检测的准确率。相比传统的安全带检测方法,深度学习的最大优势在于它可以自动的从样本数据中学习特征,最大限度的减少了人为的干预以及手工设计特征的复杂性。本方法中在利用深度学习进行训练和检测之前,先要对交通视频图像进行预处理。即首先运用帧差法获取运动车辆的最小外接矩形,然后根据经验阈值对车窗进行粗定位排除车头部分存在的干扰信息,再利用边缘检测与积分投影对车窗进行精确定位,从而获得驾驶人员所在车窗区域作为训练样本图片或检测图片。之后研究了深度学习中的卷积神经网络模型,并用卷积神经网络模型对样本图片进行训练得到用于安全带检测的模型,然后用该模型对待检测图片进行检测,判断出车辆司机是否佩带安全带。本文中利用Caffe框架对整个训练和检测过程进行了实现,并对实验结果进行了分析,证明了该方法的有效性。
[Abstract]:In recent years, the number of motor vehicles in our country has been increasing continuously, and the number of traffic accidents and casualties has been increasing year by year. As a very important passive protection measure, safety belt can effectively reduce the casualties caused by vehicle collisions or other traffic accidents while driving on the road. However, the phenomenon that drivers do not wear seat belts in the course of driving is still widespread in China. The main reason is that drivers'safety awareness is not strong, and there are many kinds of irregular behaviors to avoid the seat belt prompting system. The detection method of wearing seat belts is very important for improving drivers'awareness of compliance with traffic regulations. This paper presents a new method of seat belts detection based on depth learning, which tries to improve the accuracy of seat belts detection compared with traditional methods. Seatbelt detection method, the biggest advantage of depth learning is that it can automatically learn features from the sample data, minimize the complexity of human intervention and manual design features. The frame difference method is used to obtain the minimum outer rectangle of the moving vehicle, and then coarse positioning of the window is carried out according to the experience threshold to eliminate the interference information in the front part of the vehicle. Then the window is accurately located by edge detection and integral projection, and the driver's window area is obtained as the training sample picture or detection picture. A convolution neural network model for in-depth learning is proposed, and the convolution neural network model is used to train the sample pictures to get a model for seat belt detection. Then the model is used to detect the pictures to be detected and determine whether the driver wears a seat belt. Finally, the experimental results are analyzed, and the effectiveness of the method is proved.
【学位授予单位】：华中科技大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：U491.61;TP391.41

【相似文献】