基于空时关系学习的运动检测和目标跟踪研究

发布时间：2018-09-04 08:48

【摘要】：智慧城市是国家解决当前城市发展问题、增加新的经济增长点、抢占未来科技制高点的重要战略,其核心建设内容之一是智能交通。智能交通的关键技术大多涉及计算机视觉。本文利用空时关系学习对复杂场景下计算机视觉中运动目标检测和目标跟踪两个核心问题进行了技术探索,研究成果应用于智能交通之智能电子警察系统,提高了电子警察系统对环境的适应性。对于运动目标检测问题,分析了面对复杂场景代表性的运动检测方法设计中存在的不足,归纳出形成复杂场景的主要因素,深入分析了光照变化、背景扰动、相似目标、相机运动等因素对运动目标检测产生不利影响的机理,分别提出了综合利用视频图像序列在不同层面的多个因素、利用目标局部特征和空时关系以及利用目标与周围环境的空时置信关系等进行运动目标检测的方法；本文还对未知目标的长时间跟踪问题进行了研究。复杂场景下的未知目标长时跟踪面临的问题包括：目标遮挡、目标外观变化、目标尺度变化以及目标的短暂消失。深入分析了目标遮挡以及目标外观变化等情况造成目标特征缺失或者不完整的情况下,仍可利用的信息,分析并比较了代表性目标跟踪算法应对目标尺度变化和目标短暂消失的处理策略,提出了一种结合目标自身特征和目标与周围环境的空时联系,可以长时间对未知目标进行稳定跟踪的方法；最后,将以上研究成果应用于智能电子警察系统,解决了研发过程中遇到的技术困难。本文的主要研究成果和贡献：1.分析了视频目标检测中复杂场景的主要组成因素,提出一种基于尺度不变局部三元模式(SILTP)的视频图像背景建模算法。根据复杂场景对视频图像序列不同层次的不同影响,利用图像帧级、图像块级和像素级三级信息设计背景建模算法。算法融合图像帧、图像块和图像像素三个层面的优势来应对复杂场景。在图像帧级,利用全局灰度均值处理场景亮度突变；在图像块级,利用SILTP纹理图像基于图像块进行背景建模,快速定位前景目标大致轮廓;在像素级,用类ViBe算法检测前景目标精确边界。此算法挖掘空时信息并融合利用,其性能在标准视频集CDM’14上得到验证。2.面对视频目标检测的难点一目标自身投影的消除问题,构建了阴影光照模型,分析了目标阴影的种类及产生的原因。将纹理信息、色调信息和空时信息与ViBe算法相结合,提出了SAViBe+算法。首先,利用图像纹理对光照变化的弱敏感性,消除室内弱光照产生的目标投影；然后,在HSV颜色空间构建色调(Hue)模型,利用物体颜色的固有特性消除室外光照造成的目标投影；最后,为了加强目标投影的消除效果,同时提高处理速度,利用像素变化的局部相关性设计了MofV因子。用标准视频集CDM’14验证了该算法的性能。3.提出在HSV颜色空间实现鲁棒运动检测的方法DMSTAB。在HSV颜色空间,通过K-means聚类,利用像素集的空时关联产生像素的局部强度差,利用单高斯模型分别为像素的局部强度差和色调建模,然后,联合两者的结果寻找潜在的阴影像素点；接着,深入分析了ViBe背景差算法的工作原理,提出基于AdaBoost-Like方法利用潜在的阴影像素点构建双关联背景模型,实现对运动目标快速精确的检测,有效消除运动目标的自身投影。用标准视频集CDM’14上多种复杂场景验证了该方法的性能。4.提出基于空时置信关系进行运动目标检测的方法STR。本文提出一种空时置信关系,定义了像素点与其环境邻域像素点之间一种相对稳定的联系。首先,根据视觉聚焦特性和光照影响图像亮度变化的规律,定义像素点与环境像素点的空域关系；然后,利用快速核密度估计方法对空域关系的时域变化建模；此外,根据空域关系值的分散度为模型分配相应的权重；最后,通过基于权重的概率综合得到像素点属于背景的概率,完成运动目标检测。该算法性能在标准视频集CDM’14的典型复杂场景中得到验证。5.提出一种将目标与其环境的空时关联信息和目标自身特征结合使用,对未知目标进行长时间、稳定跟踪的新方法LST。该方法借鉴TLD算法框架,通过检测和跟踪两种独立途径对目标进行跟踪。算法包括检测、跟踪和学习三个功能模块。检测模块通过若干分类器级联,根据目标自身基本的图像特征在全局范围内检测目标,处理目标短暂消失又重现、目标尺度变化以及环境干扰；跟踪模块利用目标与其周围环境的空时置信关系,通过局部搜索,快速跟踪目标,处理目标遮挡、目标尺度变化；算法在运行过程中,通过维护一组由正样本组成的在线模板,对跟踪和检测效果进行评测。学习模块依据评测结果,调整检测模块和跟踪模块相关参数,实现算法的自学习。在若干对跟踪算法极具挑战性(严重遮挡、剧烈的光照变化、姿态和尺度变化、非刚性形变、复杂背景、运动模糊和相似目标)的数据集上比较了LST算法与主流视频目标跟踪算法的性能,LST算法展现出了较好的跟踪效果。6.面对电子警察系统研发过程中遇到的技术瓶颈,将运动目标检测算法STR和目标跟踪算法LST的核心技术应用于智能电子警察系统。提高了智能电子警察系统的车辆检测和车辆跟踪性能,并进一步作用于车牌识别和车辆违章行为评判,提高了电子警察系统的整体性能。该电子警察系统首期工程已经通过验收。
[Abstract]:Intelligent city is an important strategy for a country to solve the current urban development problems, increase new economic growth points, and seize the commanding heights of future science and technology. Intelligent traffic is one of its core construction contents. Two core problems of detection and target tracking are explored technically. The research results are applied to the intelligent electronic police system of intelligent transportation, which improves the adaptability of the electronic police system to the environment. The main factors of complex scenes are analyzed. The mechanism of unfavorable effects of illumination changes, background disturbance, similar targets, camera motion and other factors on moving target detection is analyzed in depth. The problem of long-term tracking of unknown targets in complex scenes includes: object occlusion, object appearance change, object scale change and short-term disappearance of the target. In the case of target feature missing or incomplete due to the change of target appearance, the available information is analyzed and compared. A representative target tracking algorithm is proposed to deal with the change of target scale and the short-term disappearance of target. Finally, the above research results are applied to the intelligent electronic police system to solve the technical difficulties encountered in the development process. The main research results and contributions of this paper are as follows: 1. The main components of complex scenes in video object detection are analyzed, and a scale-invariant approach is proposed. Local ternary mode (SILTP) video image background modeling algorithm. According to the different effects of complex scenes on different levels of video image sequences, the background modeling algorithm is designed using three levels of information: frame level, image block level and pixel level. The algorithm combines the advantages of image frame, image block and image pixel to deal with complex scenes. At the frame level, the global gray mean is used to deal with the sudden change of scene brightness; at the image block level, the SILTP texture image is used to model the background based on the image block to quickly locate the outline of the foreground target; at the pixel level, the precise boundary of the foreground target is detected by the ViBe-like algorithm. Confronted with the difficulty of video object detection, i.e. the elimination of object self-projection, a shadow illumination model is constructed, and the types and causes of object shadows are analyzed. The weak sensitivity of illumination changes eliminates the target projection caused by indoor weak illumination; then, a hue model is constructed in HSV color space to eliminate the target projection caused by outdoor illumination by using the intrinsic characteristics of object color; finally, in order to enhance the elimination effect of target projection and improve the processing speed, the local correlation of pixel changes is used. MofV factor is designed. The performance of the algorithm is verified by the standard video set CDM'14. 3. A robust motion detection method DMSTAB is proposed in HSV color space. In HSV color space, the local intensity difference of pixels is generated by K-means clustering, and the local intensity difference of pixels is generated by spatial-temporal correlation of pixel sets. Then, the working principle of Vibe background subtraction algorithm is deeply analyzed, and a bi-correlation background model based on AdaBoost-Like method is proposed to detect moving objects quickly and accurately and eliminate moving objects effectively. Projection. The performance of this method is validated by a variety of complex scenes on the standard video set CDM'14. 4. A space-time confidence relation based moving object detection method STR is proposed. In this paper, a space-time confidence relation is proposed, and a relatively stable relation between pixels and their neighborhood pixels is defined. Then, a fast kernel density estimation method is used to model the temporal variation of spatial relationship. In addition, the corresponding weights are assigned to the model according to the dispersion of spatial relationship values. Finally, the pixels are synthesized by the probability based on weights. The algorithm is validated in typical complex scenes of standard video set CDM'14. 5. A new method LST is proposed, which combines the space-time association information of the target and its environment with the target's own characteristics to track unknown targets for a long time and stably. The algorithm consists of three functional modules: detection, tracking and learning. The detection module cascades through several classifiers, detects the target in the global scope according to the basic image features of the target itself, handles the transient disappearance and recurrence of the target, changes in target scale and environment. The tracking module uses the space-time confidence relationship between the target and its surroundings to track the target quickly through local search, deal with the occlusion of the target and the change of the target scale; the algorithm evaluates the tracking and detection effect by maintaining a set of online templates composed of positive samples in the running process. The learning module adjusts the tracking and detection results according to the evaluation results. The LST algorithm is compared with the mainstream video target tracking algorithm on several datasets which are challenging to the tracking algorithm (severe occlusion, drastic illumination changes, attitude and scale changes, non-rigid deformation, complex background, motion blur and similar targets). ST algorithm shows a good tracking effect. 6. In the face of the technical bottleneck encountered in the development of the electronic police system, the core technology of moving target detection algorithm STR and target tracking algorithm LST is applied to the intelligent electronic police system. The vehicle detection and tracking performance of the intelligent electronic police system are improved, and further acts on it. License plate recognition and vehicle violation judgment have improved the overall performance of the electronic police system.
【学位授予单位】：西安电子科技大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.41

【相似文献】