基于CNN和多尺度融合的驾驶员打电话行为检测

Driver’s Call Behavior Detection Based on CNN and Multi-scale Fusion

由于传统的驾驶员违规接听电话行为检测方法缺乏一套严谨的评判模型,难以满足现实中驾驶员违规接听电话的识别需要,因此如何建立一套合理有效的评判模型成为亟待解决的问题。针对目前评判模型的局限性,采用计算机视觉技术和深度学习模型相结合的方式对驾驶员违规接听电话行为进行科学评判。主要是通过提取的Haar-Like特征训练级联分类器捕获脸部特征,采用CNN模型和ROI技术提取手部特征,并利用YoloV3目标检测算法识别手机,依据特征间的空间位置关系来判断驾驶员是否存在违章接听电话行为。通过大量数据的实验测试,结果证明了该评判模型不仅能将精确度提高至96.28%,而且能实时检测到行车时违规接听电话行为并进行提醒,进而降低因违规接听电话发生交通事故的概率。

Due to the lack of a set of rigorous judgment models for the traditional method of detecting illegal calls by drivers, it is difficult to meet the recognition needs of illegally answering calls in reality. Therefore, how to establish a reasonable and effective evaluation model has become an urgent problem to be solved. Aiming at the limitation of the evaluation model, computer vision technology and deep learning model are used to scientifically evaluate driver violation call. The facial features are captured by the extracted Haar-Like feature training cascade classifier, the hand features are extracted by CNN model and ROI technology, and the mobile phone is identified by YoloV3 target detection algorithm. According to the spatial location relationship between features, it can judge whether the driver has illegal phone calls behavior. Through a large number of experimental data tests, it is proved that the proposed evaluation model can not only improve the accuracy to 96.28%,but also can detect and remind the illegal phone calls while driving in real time, thus reducing the probability of traffic accidents caused by illegal phone calls.