强干扰条件下移动前景检测的层次化模型和算法

As a representative computer vision task, moving foreground detection is not only the basic problem of advanced artificial intelligence tasks like semantic understanding and driverless driving, but also widely used in practical tasks such as monitoring and early warning. In recent years, relevant research has made great progress. The difference between foreground and background becomes the core modeling basis of the relevant models. However, in practical applications, strong noise, irregular changing background and inconspicuous foreground have become strong interference factors. This project aims to explore the generation rules of the interference factors and the distribution characteristics of various video components, and accordingly proposes hierarchical joint models and algorithms which will achieve effective and stable practical performance. The main research content includes: 1) for the complex distributed noise, a structured mixture of asymmetric probability distributed noise model is proposed; 2) for the irregular changing background, by employing characterization methods such as manifold learning and deep learning, a deep hierarchical background model is proposed; 3) for the inconspicuous foreground, benefits from various analysis methods, like saliency and gradient, a multi-information fusion based foreground model is proposed; 4) build an online hierarchical joint model and the corresponding algorithm. The research of this project will explore a feasible solution for the theoretical research and practical application of moving foreground detection, and is significant theoretically and practically.

移动前景检测作为代表性的计算机视觉任务，既是语义理解、无人驾驶等人工智能研究的基础问题，又在监控、预警等实际任务中有着广泛应用。近年来，相关研究取得了长足进展。其中，前景和背景的显著差异成为各类模型的主要建模依据。但在实际应用中，强烈的噪声、不规则变动的背景和弱显著性的前景等因素往往对此形成较强的干扰。本项目针对性地探索各种强干扰因素的产生规律和视频各成分的分布特点，并据此层次化地构建更鲁棒的联合模型和算法，从而达到高效、稳定的实用效果。主要研究内容为：1）针对分布复杂的噪音，提出结构化混合非对称分布噪音模型；2）针对变动的背景，借助流形学习、深度学习等特性描述手段，提出深度层次化背景模型；3）针对不显著的前景，借助对比度、梯度等分析手段，提出多信息融合前景模型；4）构建在线的联合模型和算法。本项目的研究，为该课题的理论研究与实际应用的结合探索出一套可行方案，具有重要的理论意义和应用价值。

自主研发的智慧监控视频分析系统将是我国人工智能产业布局中的一项关键技术，而摄像头数量大、利用率低、智能化程度低是我国智能化监控视频产业的现状，针对性的智慧场景视频分析技术将促进大批量的闲置或利用效率低下的摄像头资源有效利用，助力国家建设和社会发展。本项目依次完成了以下研究内容：1）完成针对非对称混合噪声建模的推导和相关噪声分布理论构建及实验验证；并探索和推演了其他常见噪声约束模型。2）完成针对视频背景的多分辨率层次推导、多成分信息拟合和深度提取网络的调试、搭建及实验验证；并针对传统背景建模框架的非凸性推理、模糊建模等探索。3）完成针对移动前景的数据特性的探索和适应性分布拟合建模，并构建高效、稳定的移动前景提取框架；并探索并构建联合模型的快速、实时更新策略。4）针对视频数据分析挖掘过程中出现的数据聚类、视频知识推理等关联任务进行探索和建模，辅助了本项目研究的顺利进行和研究成果的有效应用。本项目的研究探索了一系列监控视频移动目标检测任务中的核心、关键技术，形成了一系列稳定性更强、运行效率更高的理论模型，大量的实验验证也分别证明了本项目一系列研究成果的价值和意义。通过本项目的研究，监控视频智能分析领域的理论模型与实际应用的联系更加紧密，部分研究成果能够直接应用于实际任务。受本项目支持，培养研究生多名，共计发表学术论文9篇，其中SCI 1区/2区及CCF A类/B类共计6篇，CCF推荐中文期刊3篇；获授权专利1项；已公开1项；申请专利多项。.。