基于SSD的眨眼与打哈欠检测

提出了一种新的基于SSD网络的眨眼与打哈欠检测.首先,对原SSD架构进行四个方面的修改.第一,主干网用MobileNetV3替换了VGG,使得架构尺寸缩减三分之一.第二,重新设计了输出检测器,实现了目标检测与分类、目标定位与高宽比回归多任务的融合.第三,网络直接回归高宽比,比传统的从关键点离线计算高宽比的方法鲁棒性要高...

SSD-based eye-blink and mouth-yawn detection

A new SSD-based eye-blink and mouth-yawn detection is proposed.At first,the original SSD architecture is modified in four aspects.Firstly,the backbone network VGG is replaced with MobileNetV3,which reduces the size of the architecture by one third.Secondly,the detector at output layer is redesigned to realize the integration of target detection with multiple detection tasks,such as classification,target localization and aspect-ratio regression for eye and mouth.Thirdly,the network directly performs the aspect ratio regression,which has higher robust than the traditional method of offline calculation of the aspect ratio from key points.Finally,the calculation of the intersection over union(IOU)and the threshold setting are dropout in the non-maximum suppression(NMS)layer,which speed up the single target detection by 40%.In addition,the performance of the proposed method is evaluated with the dataset YawDD.The obtained results show that the detection accuracy of blinking and yawning are 95.1%and 100%,respectively,with a high detection speed of 41fps achieving on a single card 2080Ti.