Li Xicai, Wu Qinqin, Xiao Bangpeng, Liu Xuanyi, Xu Chen, Li Xueling, Xu Bin, Wang Yuanqing
Appl Opt. 2020 May 10;59(14):4199-4208. doi: 10.1364/AO.386903.
In order to localize the viewers' eyes, a high-speed and robust infrared-guiding multiuser eye localization system was fabricated in this paper for a binocular autostereoscopic display, which can project a pair of parallax images to corresponding eyes. The system is composed of a low-resolution thermal infrared camera, a pair of high-resolution left and right visible spectral cameras, and an industrial computer. The infrared camera and the left visible spectral camera, and the left and right visible spectral camera, can both form the binocular vision system. The thermal infrared camera can capture the thermography images. The left and right visible spectral cameras can capture the left and right visible spectral images, respectively. Owing to the temperature difference between the face and background, the features of the face in thermography images are prominent. We use the YOLO-V3 neural network to detect the viewers' faces in thermography images. Owing to the different features of the pseudo and real faces in the infrared spectral, in the thermography images, the pseudo-faces can be easily eliminated. According to the positions and sizes of potential bounding boxes of the detected faces in the thermography images, the industrial computer can be guided to determine the left candidate regions in the left visible spectral image. Then, the industrial computer can determine the right candidate regions in the right visible spectral image. In the left candidate regions, the industrial computer detects the faces and localize the eyes by using the SeetaFace algorithm. The template matching is performed between the left and right candidate regions to calculate the accurate distance between the viewer and the system. The average detection time of the proposed method is about 3-8 ms. Compared with traditional methods, the localization time is improved by 86.7%-90.1%. Further, the proposed method is hardly influenced by the pseudo-faces and the strong ambient light.
为了定位观看者的眼睛，本文为双目自动立体显示器制作了一种高速且鲁棒的红外引导多用户眼睛定位系统，该系统可以将一对视差图像投射到相应的眼睛上。该系统由一个低分辨率热红外相机、一对高分辨率的左右可见光谱相机和一台工业计算机组成。红外相机与左可见光谱相机以及左右可见光谱相机均可构成双目视觉系统。热红外相机可以捕捉热成像图像。左右可见光谱相机可以分别捕捉左右可见光谱图像。由于面部与背景之间的温度差异，热成像图像中面部的特征较为突出。我们使用YOLO-V3神经网络在热成像图像中检测观看者的面部。由于红外光谱中伪脸和真实脸的特征不同，在热成像图像中可以很容易地消除伪脸。根据热成像图像中检测到的面部潜在边界框的位置和大小，引导工业计算机确定左可见光谱图像中的左候选区域。然后，工业计算机可以确定右可见光谱图像中的右候选区域。在左候选区域中，工业计算机使用SeetaFace算法检测面部并定位眼睛。在左右候选区域之间进行模板匹配，以计算观看者与系统之间的准确距离。所提方法的平均检测时间约为3 - 8毫秒。与传统方法相比，定位时间提高了86.7% - 90.1%。此外，所提方法几乎不受伪脸和强光环境的影响。
