基于YOLO3的人脸自动跟踪摄像机器人系统研究

YOLO3 based face tracking camera robot system

针对虚拟演播室下对主持人的持续拍摄任务,设计了一款自动跟踪摄像机器人并提出一种基于深度卷积神经网络YOLO3的可对主持人面部进行循环快速检测与跟踪的方法。针对持续拍摄任务的实时性以及人脸的单一性,在对一阶快速目标检测算法YOLO3进行模型的优化与改进的同时使用Wider-Face数据集对其进行训练,并通过在人脸监测的循环中对机器人云台的速度进行持续的调节与控制。在FDDB数据集上的测试表明改进的YOLO3算法对人脸检测的高效与准确性,而通过对人脸的跟踪实验,证实了人脸自动追踪摄像机器人系统可以快速,准确,稳定地对主持人进行持续的跟踪拍摄,且能够泛化适应如球类运动中的球等其他类别的物体。这篇文章将首先介绍虚拟演播室系统,然后回顾各类人脸检测算法,并介绍以目标检测算法YOLO3为基础而训练的人脸跟踪算法,以及基于此算法的人脸自动跟踪摄像机器人系统。

Aiming at the continuous shooting task of the host under the virtual studio, a face automatic tracking camera robot is de- signed and a method based on YOLO3, a deep convolutional neural network, is proposed to detect and track the face of the host. In view- of the real - time performance of the continuous shooting task and the simpleness of the human face, ＂One - Stage＂ fast target detection algorithm YOLO3 is trained on Wider - Face dataset, and the speed of the robot＇ s Pan - Tilt is continuously adjus- ted and controlled in the face tracking cycle. The test on the FDDB dataset show-s the efficiency and accuracy of the modified YO- LO3 algorithm for face detection, and through the experiment of face tracking, it is proved that the automatic tracking camera sys- tem can quickly, accurately and steadily track the host. The system also enjoys a high generality and can be adapted to track balls in telecast of ball games or some other tracking tasks. This article will first introduce the virtual studio system, and then review all kinds of face detection algorithms, and introduce the face tracking algorithm based on the target detection algorithm YOLO3, and the automatic face tracking robot system based on this algorithm.