感兴趣运动目标检测的研究与实现

发布时间：2018-07-31 20:50

【摘要】：目标检测是图像视频分析、模式识别和计算机视觉应用的一个首要问题,在各个方面起着重要的作用,如视频监控、车辆导航、机器人视觉和智能交通系统等。一般查看视频的时候,人们通常只会对特定的运动目标有很大兴趣,可将其称为感兴趣目标。感兴趣目标的准确快速提取将大大提高后续跟踪和识别处理的有效性。在实际应用中,面对海量视频图像以及不同应用场合,需要解决感兴趣运动目标检测的准确性、实时性和平台通用性问题。具体研究工作如下:假设视频中的运动目标都是感兴趣目标,基于能量泛函的图像分割是广泛使用的方法之一。针对现今基于能量泛函的图像分割计算效率不高且平台局限的特点,首先提出了连续最大流图像分割算法在Open CL上跨平台并行实现。在分析连续最大流算法的并行特征基础上,将迭代求解最大流优化问题并行实现。合理调用异构平台的CPU和GPU,使算法具有高性能和可移植性。然后将所提算法应用于视频,在混合高斯背景模型下,结合并行后的连续最大流图像分割算法进行感兴趣运动目标检测。实验结果表明,在保证视频图像的分割质量下,算法的GPU+CPU异构并行实现较CPU串行实现有数量级的提升;算法在AMD,Nvidia和Intel三大主流硬件平台上通用运行,验证了算法的有效性和平台的可移植性,基本满足实际应用的要求。实际视频分析中并非对所有运动目标都感兴趣,需要根据不同的应用场合提取某一类特定的感兴趣运动目标。针对由于复杂背景和其他非感兴趣运动目标的干扰而导致不能准确检测出感兴趣运动目标的问题,本文提出了一个新的感兴趣运动目标检测框架。在传统的马尔科夫随机场(Markov Random Field,MRF)基础上,引入haar-like级联分类器搜索到的感兴趣建议区域作为高阶势能项,从而构建高阶MRF模型。在这统一能量框架下,通过降阶优化最终转为用最大流/最小割解决能量最优化问题。实验结果表明,引入haar-like特征训练的级联分类器搜索得到的感兴趣建议区域作为高阶势能项进行建模,提高了模型的表达能力;同时算法有效增强了感兴趣目标的分割效果,尤其是使感兴趣目标有更好的边缘性,提高分割的准确度,改善分割感兴趣目标的视觉效果。
[Abstract]:Target detection is one of the most important problems in image video analysis, pattern recognition and computer vision applications. It plays an important role in many aspects, such as video surveillance, vehicle navigation, robot vision and intelligent transportation system. When viewing a video, people usually only have a great interest in a particular moving object, which can be called the object of interest. Accurate and fast extraction of objects of interest will greatly improve the effectiveness of follow-up tracking and recognition processing. In practical applications, in the face of mass video images and different applications, it is necessary to solve the problems of accuracy, real-time and platform generality of moving object detection of interest. The research work is as follows: assuming that the moving targets in video are all objects of interest, image segmentation based on energy functional is one of the widely used methods. In view of the low efficiency and limited platform of image segmentation based on energy functional, a continuous maximum flow image segmentation algorithm is proposed in this paper, which is implemented in parallel on Open CL platform. Based on the analysis of the parallel characteristics of the continuous maximum flow algorithm, the iterative solution to the maximum flow optimization problem is implemented in parallel. The algorithm has high performance and portability by reasonably calling CPU and GPU of heterogeneous platform. Then, the proposed algorithm is applied to video, and combined with parallel continuous maximum flow image segmentation algorithm in hybrid Gao Si background model to detect moving objects of interest. The experimental results show that the GPU CPU isomerous parallel implementation of the algorithm is of the order of magnitude higher than that of the CPU serial implementation, and the algorithm is running on the three main hardware platforms of Intel and AMD-Nvidia under the guarantee of video image segmentation quality. The validity of the algorithm and the portability of the platform are verified. Not all the moving targets are interested in the actual video analysis, so it is necessary to extract a certain kind of moving objects according to different applications. In view of the problem that the moving objects of interest can not be detected accurately due to the interference of complex background and other moving objects of non-interest, a new frame for detecting moving objects of interest is proposed in this paper. Based on the traditional Markov Random Field (Markov Random), the region of interest found by the haar-like cascade classifier is introduced as a high-order potential energy term, and the higher-order MRF model is constructed. In this unified energy framework, the reduced order optimization is eventually converted to the maximum flow / minimum cut to solve the energy optimization problem. The experimental results show that the proposed region of interest obtained by cascaded classifier based on haar-like feature training is modeled as a high-order potential energy item, and the expression ability of the model is improved, and the segmentation effect of the object of interest is effectively enhanced by the algorithm. Especially, it can improve the edge of the object of interest, improve the accuracy of segmentation and improve the visual effect of the object of interest.
【学位授予单位】：杭州电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41

【参考文献】