Mobile face detection and tracking for media streaming applications
Mobile face detection and tracking for media streaming applications
复制标题
DOI：
10.1504/ijwmc.2007.014457
复制
发表时间：
2007-07
期刊：
影响因子：
--
通讯作者：
D. Tsishkou;Liming Chen;E. Bovbel
中科院分区：
文献类型：
--
作者：
D. Tsishkou;Liming Chen;E. Bovbel
本文提出了一种用于流媒体应用中的移动的人脸检测和跟踪的方法。为了说明移动的图像处理和模式识别的具体情况，对Intel x86和Nokia S60平台进行了实验基准测试。为了记录关于移动的媒体流用户的简档，考虑了关于移动的电话使用的各个方面。该配置文件是通过分析移动的手机用户的照片和他们的设备截图生成的。移动的人脸检测模型基于鲁棒实时对象检测算法的优化版本。我们的优化能够将原始版本的计算复杂度降低4倍。我们还描述了跟踪模式，其中系统实现了近实时的帧速率（45 - 12 fps，取决于视频的复杂性）。人脸检测器的训练是在超过400小时的视频数据库上进行的，主要参数如FRR和FAR分别在30分钟的视频数据库（包括超过5000张人脸图像）和200小时的视频数据库上进行评估。人脸跟踪基准使用20个移动的视频序列，其中涵盖了广泛的人类活动和环境变化。我们描述了低比特率视频编码器的结构，该编码器使用人脸检测和跟踪来分割前景/背景对象，并以不同的比特率对它们进行编码。我们还描述了一个框架，智能视频会议和低比特率图像编码的实验结果，为这个应用程序。
This paper presents a method for mobile face detection and tracking in media streaming applications. To account specifics of mobile image processing and pattern recognition, the experimental benchmarking of Intel x86 and Nokia S60 platforms is done. Various aspects on mobile phone use are considered in order to record a profile on a mobile media streaming user. The profile is generated by analysing pictures of mobile phone users and screenshots of their devices. The mobile face detection model is based on the optimised version of robust real-time object detection algorithm. Our optimisation enables to reduce the computational complexity of the original version in 4 times. We also describe the tracking mode in which the system achieves a near real-time frame rate (45 12 fps depending on the video complexity). The training of the face detector is done on the databases of more than 400 hr of video and the major parameters such as FRR and FAR are evaluated on a database of 30 min of video (including more than 5000 facial images) and 200 hr of video, respectively. Face tracking is benchmarked using 20 mobile video sequences, which cover a broad range of human activities and environmental changes. We describe the construction of the low bit-rate video coder that uses face detection and tracking to segment foreground/background object and encode them with a different bit-rate. We also describe a framework for smart videoconferences and experimental results on low bit-rate image coding for this application.