彩色/多光谱异源双目视频运动目标分割方法研究

项目名称： 彩色/多光谱异源双目视频运动目标分割方法研究

项目编号： No.61501456

项目类型： 青年科学基金项目

立项/批准年度： 2016

项目学科： 无线电电子学、电信技术

项目作者： 张耿

作者单位： 中国科学院西安光学精密机械研究所

项目金额： 21万元

中文摘要： 视频运动目标分割是视频分析领域中一项极富挑战性的课题。在彩色视频中，当目标与背景颜色相似时，边缘提取、运动估计和深度估计都会面临困难。本项目将多光谱视频的高区分度与彩色视频的高分辨率有机结合，提出一种基于彩色与多光谱视频的异源双目视频运动目标分割方法。在申请人已有的轨迹跟踪和分割模型的基础上，参考现有的多光谱图像分割和配准方法，本项目深入讨论不同分辨率彩色和多光谱视频的精确配准问题，并提出基于异源双目视频同时估算运动场及深度场的方法；在此基础上，进一步研究三维点轨迹和区域轨迹的生成方法；最后，利用多层图模型融合像素光谱、点轨迹和区域轨迹三种不同粒度特征并优化求解。本项目所研究的算法利用了多光谱视频与彩色视频的互补特性，能够准确估计深度场及运动场；所提模型融合多粒度特征，提高了复杂场景下的运动目标分割精度。相应的研究成果可以为智能监控、自主瞄准、辅助驾驶等领域提供前端支持。

中文关键词： 运动识别；目标检测；立体视觉；多源数据融合

英文摘要： Moving object segmentation in videos is a challenging problem in the field of computer vision. In color videos, when the object has similar color with the background, it raises problems in boundary extraction, motion estimation and depth estimation. This project integrates the distinguishing ability of multispectral video and the high resolution of color video, and proposes a moving object segmentation algorithm using heterogeneous binocular video. Based on the applicant’s previous work on trajectory tracking and segmentation model, and considering the current work on multispectral image segmentation and registration algorithms, this project discusses the problem of registering color image and multispectral image under different resolutions, and proposes a method to simultaneously estimate motion and depth based on heterogeneous binocular video. Furthermore, this project explores methods to generate three dimensional point trajectories and region trajectories. Finally, features of different granularities, including pixel spectral, point trajectories and region trajectories, are integrated into a multi-layer graph model and solved for object segmentation. The proposed algorithm utilizes the complementary properties of multispectral video and color video, and is able to estimate depth and motion accurately. The proposed model integrates multi-granular features and increases segmentation accuracy. The research achievements can be used in fields of smart surveillance, self-aiming, and driving assistance.

英文关键词： Motion Recognition;Object Segmentation;Stereo Vision;Multi-source Data Fusion