基于YOLOv3和RGBD的车内人耳识别与深度定位  

On-board Human Ear Recognition and Depth Coordinates Location Based on YOLOv3 and RGBD

作　　者：金银强 王岩松 张伟伟 王孝兰 JIN Yinqiang;WANG Yansong;ZHANG Weiwei;WANG Xiaolan(College of Mechanical&Automotive Engineering,Shanghai University of Engineering Science,Shanghai 201620)

机构地区：[1]上海工程技术大学机械与汽车工程学院,上海201620

出　　处：《计算机与数字工程》2021年第10期2096-2101,共6页Computer & Digital Engineering

基　　金：国家自然科学基金项目(编号:51675324)资助。

摘　　要：定位人耳空间深度坐标是车内主动降噪研究的关键环节。论文基于YOLOv3深度卷积网络的目标检测算法对乘员耳部进行检测,并结合RGBD多源视觉传感系统动态定位车内人耳深度坐标。首先,采集车内乘员侧面RGB图像作为人耳数据集,并在人耳区域进行标定。然后结合YOLOv3的Darknet-53网络训练得到人耳检测模型。最后,融合RGBD视觉传感系统,完成人耳识别与深度坐标定位系统。实验表明,该方法在多种光照下,对车内人耳动态跟踪速度达到7fps,检测准确率在95%以上。同时,空间深度定位精度达到±3.5cm,可为车内主动降噪系统提供准确的噪声信号定位反馈。Locating spatial depth coordinate of human ear is inevitable to active noise reduction in vehicle.This paper presents a procedure for dynamical locating spatial depth coordinate of human ear with the RGBD multi-source visual sensing system that is combined with target detection algorithm of YOLOv3 deep convolution network under the premise of successful detection of the occupant's ear.In order to set up a human ear recognition and depth coordinate locating system,human ear detection modal is established in the calibrated values within the occupant's ear that is collected from the profile RGB images are trained in the Darknet-53 network of YOLOv3.Experiment results show that dynamic tracking speed of the human ear in the vehicle can reach 7fps in various illuminations,and the detection accuracy is over 95%with this method.At the same time,the spatial locating accuracy reaches±3.5cm,which can provide accurate depth location of human ear for the active noise reduction system of vehicle.

关 键 词：主动降噪 深度卷积网络 YOLOv3 人耳检测 深度坐标定位 

分 类 号：TN219[电子电信—物理电子学]