基于背景建模的运动目标监控视频检测算法
发布时间：2018-07-20 20:51
【摘要】：视频监控已广泛应用于水路航运,尤其对运动船舶尺寸、流量、速度以及异常事件分析的需求日益突出,运动目标检测作为视频监控系统的核心环节,作用不言而喻。然而实际监控场景复杂多变,变化的光线、摇晃的树叶、水面的波纹等,对运动目标检测的准确性产生了巨大影响,有必要专题研究复杂场景下的运动目标检测算法。论文源自船联网国家重大专项《船舶实时视频图像监测识别系统关键技术研究及应用》和江苏省海事局科研项目《船舶超限检测系统》,结合项目中遇到的难点和挑战,诸如水面波纹、船尾拖纹、相机抖动、船舶运动速度过慢等问题,分析了国内外现有的背景建模算法,找出了算法误检的原因,针对算法缺陷提出了优化方案,并加以实验验证。论文阐述了国内外视频监控系统的研究现状,解析了背景建模算法在运动目标检测中的作用和地位,对比分析了经典背景建模算法的检测性能,提出了基于改进VIBE的运动目标检测算法,优化了基于像素运动特征的MOG算法,给出了基于三维图像的运动船舶超限监测算法。论文研究了背景建模算法理论,对比分析了六种经典背景建模算法的优势与不足,通过实验验证了各个算法的适用场景和检测效果。论文分析了VIBE算法的优点与缺陷,针对动态背景问题,从背景模型初始化、模型匹配和模型更新三个阶段,采用多帧连续图像初始化背景模型,削弱了“鬼影”对后续检测的干扰；根据背景动态程度自适应调整匹配阈值,减少了动态背景误检；计算背景样本离散度寻找最优替换样本,结合空间一致性原理和模糊理论,提高了背景模型准确性,降低了误检率。论文分析了相机抖动下背景像素的运动信息分布,提取像素运动特征,二次分辨MOG算法提取的运动前景,剔除相机抖动产生的误检,并自适应优化了算法的学习速率,在前景区域和背景区域设置不同的更新速率,解决了动态背景误检和真实前景漏检之间的矛盾,提高了算法的鲁棒性。论文研究了三维图像运动船舶超限监测算法,采用三台扫描仪采集三维点云,运用PCA方法分类点云数据,测量出运动船舶的三维尺寸。论文创新点如下：●提出基于改进VIBE的运动目标检测算法,采用多帧图像初始化背景模型,根据背景动态程度自适应调整匹配阈值,计算样本离散度实现最佳样本替换,结合模糊理论,提高了背景模型准确性,降低了动态背景对运动目标检测的干扰；·优化基于像素运动特征的MOG算法,提取像素运动特征用于运动前景二次分辨,并自适应调整学习速率,抑制了抖动误检,提高了算法的鲁棒性。·给出了基于三维图像的运动船舶超限监测算法,用于检测运动船舶的三维尺寸,为监测超限船舶提供依据。
[Abstract]:Video surveillance has been widely used in waterway shipping, especially for the size, flow, speed and abnormal event analysis of moving ships. As the core of video surveillance system, the role of moving target detection is self-evident. However, the complex and changeable scene, changing light, rocking leaves and ripples of water surface have a great influence on the accuracy of moving target detection. It is necessary to study the algorithm of moving target detection in complex scene. The thesis is derived from the key technology research and application of ship real-time video image monitoring and identification system, a major national project of ship networking, and the scientific research project of Jiangsu Marine Administration, "ship Over-limit Detection system", which combines the difficulties and challenges encountered in the project. Problems such as surface ripple, stern ripple, camera jitter and slow ship motion are analyzed. The existing background modeling algorithms at home and abroad are analyzed, and the reasons for the false detection of the algorithm are found out, and the optimization scheme is proposed for the defects of the algorithm. It is verified by experiments. This paper describes the research status of video surveillance system at home and abroad, analyzes the role and status of background modeling algorithm in moving target detection, and compares the detection performance of classical background modeling algorithm. A moving target detection algorithm based on improved vibe is proposed, and a Mog algorithm based on pixel motion features is optimized. This paper studies the theory of background modeling algorithm, compares and analyzes the advantages and disadvantages of six classical background modeling algorithms, and verifies the applicable scene and detection effect of each algorithm through experiments. This paper analyzes the advantages and disadvantages of VIBE algorithm. Aiming at the dynamic background problem, the background model is initialized from three stages: background model initialization, model matching and model updating, and multi-frame continuous image is used to initialize the background model. It weakens the interference of "ghost" to the subsequent detection; adaptively adjusts the matching threshold according to the dynamic degree of background, reduces the false detection of dynamic background; calculates the dispersion of background samples to find the optimal replacement samples, and combines the spatial consistency principle and fuzzy theory. The accuracy of the background model is improved and the false detection rate is reduced. In this paper, the motion information distribution of background pixels under camera jitter is analyzed, the motion feature of pixels is extracted, the motion foreground is extracted by second resolution Mog algorithm, the false detection caused by camera jitter is eliminated, and the learning rate of the algorithm is optimized adaptively. Different update rates are set in the foreground region and background area, which solves the contradiction between dynamic background false detection and real foreground missed detection, and improves the robustness of the algorithm. In this paper, we study the algorithm of detecting the moving ship in 3D image. Three scanners are used to collect the 3D point cloud, and PCA method is used to classify the point cloud data to measure the 3D size of the moving ship. The innovations of this paper are as follows: a moving target detection algorithm based on improved vibe is proposed. The background model is initialized by multi-frame images and the matching threshold is adjusted adaptively according to the dynamic degree of background. The sample dispersion is calculated to achieve the best sample replacement. Combined with fuzzy theory, the accuracy of background model is improved, the interference of dynamic background to moving target detection is reduced, and the Mog algorithm based on pixel motion feature is optimized to extract pixel motion feature for the second resolution of motion foreground. It adaptively adjusts the learning rate, restrains the jitter error detection, and improves the robustness of the algorithm. An algorithm based on 3D image is presented to detect the 3D size of the moving ship, which provides the basis for monitoring the ship in excess of the limit.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41;TN948.6
本文编号：2134739
[Abstract]:Video surveillance has been widely used in waterway shipping, especially for the size, flow, speed and abnormal event analysis of moving ships. As the core of video surveillance system, the role of moving target detection is self-evident. However, the complex and changeable scene, changing light, rocking leaves and ripples of water surface have a great influence on the accuracy of moving target detection. It is necessary to study the algorithm of moving target detection in complex scene. The thesis is derived from the key technology research and application of ship real-time video image monitoring and identification system, a major national project of ship networking, and the scientific research project of Jiangsu Marine Administration, "ship Over-limit Detection system", which combines the difficulties and challenges encountered in the project. Problems such as surface ripple, stern ripple, camera jitter and slow ship motion are analyzed. The existing background modeling algorithms at home and abroad are analyzed, and the reasons for the false detection of the algorithm are found out, and the optimization scheme is proposed for the defects of the algorithm. It is verified by experiments. This paper describes the research status of video surveillance system at home and abroad, analyzes the role and status of background modeling algorithm in moving target detection, and compares the detection performance of classical background modeling algorithm. A moving target detection algorithm based on improved vibe is proposed, and a Mog algorithm based on pixel motion features is optimized. This paper studies the theory of background modeling algorithm, compares and analyzes the advantages and disadvantages of six classical background modeling algorithms, and verifies the applicable scene and detection effect of each algorithm through experiments. This paper analyzes the advantages and disadvantages of VIBE algorithm. Aiming at the dynamic background problem, the background model is initialized from three stages: background model initialization, model matching and model updating, and multi-frame continuous image is used to initialize the background model. It weakens the interference of "ghost" to the subsequent detection; adaptively adjusts the matching threshold according to the dynamic degree of background, reduces the false detection of dynamic background; calculates the dispersion of background samples to find the optimal replacement samples, and combines the spatial consistency principle and fuzzy theory. The accuracy of the background model is improved and the false detection rate is reduced. In this paper, the motion information distribution of background pixels under camera jitter is analyzed, the motion feature of pixels is extracted, the motion foreground is extracted by second resolution Mog algorithm, the false detection caused by camera jitter is eliminated, and the learning rate of the algorithm is optimized adaptively. Different update rates are set in the foreground region and background area, which solves the contradiction between dynamic background false detection and real foreground missed detection, and improves the robustness of the algorithm. In this paper, we study the algorithm of detecting the moving ship in 3D image. Three scanners are used to collect the 3D point cloud, and PCA method is used to classify the point cloud data to measure the 3D size of the moving ship. The innovations of this paper are as follows: a moving target detection algorithm based on improved vibe is proposed. The background model is initialized by multi-frame images and the matching threshold is adjusted adaptively according to the dynamic degree of background. The sample dispersion is calculated to achieve the best sample replacement. Combined with fuzzy theory, the accuracy of background model is improved, the interference of dynamic background to moving target detection is reduced, and the Mog algorithm based on pixel motion feature is optimized to extract pixel motion feature for the second resolution of motion foreground. It adaptively adjusts the learning rate, restrains the jitter error detection, and improves the robustness of the algorithm. An algorithm based on 3D image is presented to detect the 3D size of the moving ship, which provides the basis for monitoring the ship in excess of the limit.
【学位授予单位】：南京大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41;TN948.6
【共引文献】
相关期刊论文 前2条
1 周晓;赵锋;朱艳林;;基于ViBe的室外动态背景闪烁像素噪声消除方法[J];计算机应用;2015年06期
2 何广达;王洪顺;李强;;基于计算机视觉的激光检测系统研究[J];激光杂志;2015年09期
相关博士学位论文 前1条
1 毕国玲;智能视频监控系统中若干关键技术研究[D];中国科学院研究生院(长春光学精密机械与物理研究所);2015年
，本文编号：2134739
本文链接：https://www.wllwen.com/kejilunwen/wltx/2134739.html
