降雨视频分析及雨天运动目标检测关键算法研究

Current researches about moving object detection in videos always consider all possible interferences that may come along in the scenes in general and there is little research for a special category of weather condition. Considering interferences caused by rainy weather including moving raindrops and reflection shadows, starting from the physical properties and video features of rainy conditions, a moving object detection algorithm that is properly adjusted to rainy weather and immune to interferences from raindrops and reflection shadows is proposed in this project. The main research contents include: 1) False detection information that is always neglected by existing moving detection algorithms is inventively used, dynamic background categories are recognized by a secondary classification algorithm, and a moving foreground detection method specially adjusted for rainy conditions is established. 2) Considering that most of recent rain detection and removal algorithms only work in scenes that is static or with slow motion, image information rules for moving raindrops are discovered, a systematic model for raindrops imaging is established, and rain detection algorithm suitable for dynamic scenes and rain removal algorithm without perturbing interesting objects are suggested. 3) Recently, most attention has been paid on achromatic shadows detection under sunlight weather conditions. In this project, physical model of chromatic reflection shadows under rainy weather conditions is studied, photometric change rules before and after reflection shadow coverage are discovered, and a reflection shadow recognizing algorithm for rainy weather conditions is proposed.

现有关于视频运动目标检测的研究，往往笼统考虑场景中可能出现的干扰因素，缺乏专门针对某类气象场景的研究。针对降雨这类天气干扰，包括运动雨滴和雨天倒影等因素，本项目从雨这一特殊气象的物理特性和视频特征出发，构建能够适应于降雨天气的、避免雨滴和倒影干扰的视频运动目标检测算法。研究内容包括：1) 创新的利用现有运动检测算法中被忽视的误检信息，通过二次分类识别动态背景类型，构建雨天运动前景检测算法。2)针对现有视频雨滴检测与消除算法大多只适应于静态背景或仅存在缓慢运动的情况，揭示运动雨滴成像规律，构建雨滴成像系统模型，提出动态场景下雨滴检测以及不干扰兴趣目标的雨滴消除算法。3)针对现有研究大多关注日照天气下无色阴影的情况，研究雨天有色倒影形成的物理模型，揭示倒影覆盖前后图像的光度变化规律，提出雨天倒影识别算法。

本项目针对降雨这类气象环境对视频目标检测的干扰，研究了雨滴成像建模及视频去雨、视频雨滴检测及雨量估算、雨天动态背景下的目标检测、雨天倒影检测等具体关键技术：.1)研究了动态雨滴光度模型、静态雨滴光度模型、散焦雨滴光度模型，分析雨滴可见性与摄像参数的关系，揭示了雨滴可见性受到曝光时间、光圈大小和焦平面距离的影响，提出针对不同场景的需求，采取相应的能够减弱雨滴对视频图像影响的摄像参数设置方法；.2)研究了基于特征概率分布的雨滴检测方法，构建了基于亮度差和色彩张量响应特征的散聚焦雨滴判别方法，基于雨滴谱分布模型，采用伽马模型拟合观测雨滴谱，提出一种基于视频的降雨量测量方法；.3)提出一种新颖的利用误检信息消除雨滴干扰的方法，解决雨滴等动态背景严重影响目标检测和识别的问题。对像素分类的二值序列统计码元跳变次数和高电平时间作为特征向量，结合K均值方法进行二次分类，提取真实运动目标。.4)提出基于彩色对数图像处理和结构森林边缘检测进行运动倒影识别的方法。.项目执行期间，发表学术论文15篇，其中EI/SCI期刊论文4篇，EI会议论文2篇；项目执行期内，获国家发明专利授权1项，申请国家发明专利4项。