面向安防监控的双目立体视频关注度分析与编码技术研究

Under the serious social security problems, the traditional surveillance system composed of single camera cannot obtain the depth information of the target and meet the demand of criminal investigation and evidence collection. Whereas by capturing the scene depth information and helping screen the key suspect points, the binocular stereo video surveillance has aroused widespread interest. In order to solve the enormous pressure of the stereo video surveillance mass data storage and transmission,the human visual system is introduced into the stereo video compressing coding by researchers, but the perception mechanism of inter-view and the inherent characteristics of monitoring the scene are neglected, hence the coding efficiency is affected. Therefore, this application puts forward the surveillance-oriented binocular stereo video attention model and its application for video coding. Firstly, by analyzing the binocular competition perception mechanisam of stereo video, it builds up an attention model of binocular stereo video coding framework to guide its perception rate allaction. Secondly, the background temporal slowly varing featurs of the monitoring scene is introduced into the view synthesis prediction to improve the efficiency of inter-view prediction. Finally, it studies the residual filter method based on the threshold of the binocular stereo just noticeable differences to improve the efficiency of binocular stereo coding through analysing the binocular stereo video attention and its influence on the masking affect. It is expected that in the case of the same subjective quality, this research can reduce 20% of the coding bitrates to meet the needs of the traditional video surveillance system upgrading to the stereo video surveillance.

面对日益严峻的社会治安形势，传统单摄像机监控体系无法获取目标的深度信息，难以满足刑事视频侦查的需要。双目立体视频监控能捕捉场景的深度信息，帮助甄别关键嫌疑要点，引起了广泛关注。针对双目立体视频监控的海量数据对存储和传输的巨大压力，研究者将人眼视觉特性引入立体视频压缩编码，但忽略了视点间感知机理及监控场景特性，制约了编码效率的提升。为此，本申请研究面向安防监控的双目立体视频关注度模型与编码技术。首先，从立体视频的双目竞争感知机理出发，建立双目立体视频关注度模型，指导双目立体视频的感知码率分配；其次，将监控场景背景时域缓变特性引入到视点合成预测，提高视间预测效率；最后，分析双目感知机理以及关注度因素对掩蔽效应的影响，研究基于双目立体视频恰可感知失真阈值的残差滤波方法，提高双目立体监控视频编码效率。预期本研究在同等主观质量下，降低20%的编码码率，满足传统视频监控体系向立体监控升级换代的需求。

针对双目立体视频监控的海量数据对存储和传输的巨大压力，我们将人眼视觉特性、视点间感知机理及监控场景特性引入立体视频压缩编码，研究了面向安防监控的双目立体视频关注度模型及其相关编码技术，提升了立体视频编码效率。主要完成的研究工作如下：.1．建立了双目立体视频关注度模型，研究了关注区域质量优先的双目立体视频感知码率分配方法，解决了因关注度区域提取不准确而造成的码率资源分配不合理的问题，进一步优化了码率分配；.2．研究了立体视频视点合成预测编码，提出了基于自适应学习的视点合成预测多视点视频编码技术，改善了合成的视点对视点间光照的差异、深度信息误差的鲁棒性问题；提出了立体视频多视点视频间跨时空预测方法，建立了面向知识的对象表达模型的预测方法，为多源视频间的跨时空预测提供了新的思路；.3．建立了融入立体视频深度对比度因素的立体视频感知JND模型，提出了基于彩色图像相关特征的深度图滤波算法，提高了对于高水平噪声的鲁棒性，并降低了深度图的边缘突变性造成的非同类邻近像素的干扰影响；提出了基于结构张量环路中值滤波的深度视频编码技术，将计算滤波系数的约束从传统的纹理相似性约束扩展到纹理结构性约束，有效提升了深度视频的编码效率及视点合成的质量；.4．提出了相似约束稀疏表示的深度图空洞填充技术，使得空洞填充后的深度图的边缘质量更加准确且与对应彩色图的边缘更加一致；提出了基于深度图像相关特征的合成虚拟视点图像空洞修复算法，建立融合场景深度信息与纹理细节的纹理合成能量模型，增强修复内容与邻近内容的纹理一致性，提高虚拟视点图像质量。