面向广域高空监视视频的超多运动目标检测和跟踪研究

Due to very large image size, low frame rate, gray-scale pxiels, low spatial resolution, sheer number of moving objects, and the moving platform, processing and analyzing the wide area airborne surveillance imagery poses unique challenges. The purpose of this project is the development of effective moving object detection and tracking methods on the wide area airborne surveillance image data. To this end, we intend to conduct research in the following three subjects: (1) In the research of moving object detection, we employ a framework of combining a simple background subtraction and the frame difference methods. We propose a method for segmenting moving foreground and for removing parallax motion, using three frame difference images, where two are from background subtraction and one is the frame difference between the corresponding consecutive two frames. In order to increase the accuracy of moving object detection, we present a local motion model based image alignment algorithm, and a parallax removing method based on pixel correspondence construction. (2) In order to handle the problem of large number of moving object tracking, we propose to use an improved multiple hypothesis tracking framework, for which we present models for handling complex cases such as occlusions, merged and splited detections, as well as moving direction sudden change situations.（3）In order to process the large size imagery efficiently, we present methods to effectively implement the proposed methods and algorithms on GPGPU.

广域高空监视视频图像数据一般具有大尺寸、低帧频、单通道、低空间分辨率、超多运动目标、受飞行平台运动影响等特点，这使得对其进行处理和分析带来了诸多新的难题。本项目旨在研究具备能从广域高空监视图像数据中有效检测、跟踪和提取出超多运动目标体及其轨迹的方法。本项目拟展开如下研究：（1）在超多运动目标检测方面，采用基于减背景和帧差法相结合框架，提出基于三差值图像的运动分割和视差去除方法，为提高检测准确度，提出基于局部变换的配准和基于像素对应搜索的视差去除方法；（2）在大数量目标数据关联跟踪方面，采用近似多假设跟踪框架，提出基于由四种概率模型组成的假设关联后验概率模型，包括自身运动模型、领近互动模型、群体影响模型及场境约束模型，进而对目标合并、部分遮挡及转向等问题进行有效处理；（3）针对大尺寸图像计算量大问题，提出基于GPGPU架构加速计算方案, 研究在GPGPU体系下实现算法的并行计算。

广域高空监视视频图像数据一般具有大尺寸、低帧频、单通道、低物体空间分辨率、超多运动目标体、受飞行平台运动影响等特点，这使得对其进行处理和分析带来了诸多新的挑战。本项目研究了具备能从广域高空监视图像数据中有效检测、跟踪和提取出超多运动目标体及其轨迹的方法。本项目展开了如下研究：（1）在超多目标检测方面，采用基于减背景和帧差法相结合框架，提出了基于三差值图像的运动前景分割和视差去除方法，为提高检测准确度，提出基于局部变换的配准和基于像素对应搜索的视差去除方法；（2）在大数量目标数据关联跟踪方面，采用近似多假设跟踪框架，提出了基于由四种概率模型组成的假设关联后验概率模型，包括自身运动模型、领近互动模型、群体影响模型及场境约束模型，进而对目标合并、部分遮挡及转向等问题进行有效处理；（3）针对大尺寸图像计算量大问题，提出了基于GPGPU架构加速计算方案, 将研究在普通计算机GPGPU体系下实现相关方法和算法并行计算；（4）为研制亿像素广域监视摄像机，研究了相关ISP算法、图像增强算法以及图像模糊度等质量评估方法，提出了一系列具有实用价值、性能优越的算法。

内容获取失败，请点击重试